encodemb.c 55.3 KB
Newer Older
Jingning Han's avatar
Jingning Han committed
1
/*
Yaowu Xu's avatar
Yaowu Xu committed
2
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
Jingning Han's avatar
Jingning Han committed
3
 *
Yaowu Xu's avatar
Yaowu Xu committed
4 5 6 7 8 9
 * 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.
Jingning Han's avatar
Jingning Han committed
10 11
 */

Yaowu Xu's avatar
Yaowu Xu committed
12 13 14
#include "./av1_rtcd.h"
#include "./aom_config.h"
#include "./aom_dsp_rtcd.h"
Jingning Han's avatar
Jingning Han committed
15

16
#include "aom_dsp/bitwriter.h"
17
#include "aom_dsp/quantize.h"
Yaowu Xu's avatar
Yaowu Xu committed
18
#include "aom_mem/aom_mem.h"
19
#include "aom_ports/mem.h"
Jingning Han's avatar
Jingning Han committed
20

21 22 23 24
#include "av1/common/idct.h"
#include "av1/common/reconinter.h"
#include "av1/common/reconintra.h"
#include "av1/common/scan.h"
Jingning Han's avatar
Jingning Han committed
25

26
#include "av1/encoder/av1_quantize.h"
27
#include "av1/encoder/encodemb.h"
28 29 30
#if CONFIG_LV_MAP
#include "av1/encoder/encodetxb.h"
#endif
31 32 33
#include "av1/encoder/hybrid_fwd_txfm.h"
#include "av1/encoder/rd.h"
#include "av1/encoder/tokenize.h"
Jingning Han's avatar
Jingning Han committed
34

35 36 37 38 39 40
#if CONFIG_PVQ
#include "av1/encoder/encint.h"
#include "av1/common/partition.h"
#include "av1/encoder/pvq_encoder.h"
#endif

41 42 43 44
#if CONFIG_CFL
#include "av1/common/cfl.h"
#endif

45 46 47
// Check if one needs to use c version subtraction.
static int check_subtract_block_size(int w, int h) { return w < 4 || h < 4; }

Angie Chiang's avatar
Angie Chiang committed
48 49 50 51
static void subtract_block(const MACROBLOCKD *xd, int rows, int cols,
                           int16_t *diff, ptrdiff_t diff_stride,
                           const uint8_t *src8, ptrdiff_t src_stride,
                           const uint8_t *pred8, ptrdiff_t pred_stride) {
52
#if !CONFIG_HIGHBITDEPTH
Angie Chiang's avatar
Angie Chiang committed
53 54
  (void)xd;
#endif
Jingning Han's avatar
Jingning Han committed
55

Angie Chiang's avatar
Angie Chiang committed
56
  if (check_subtract_block_size(rows, cols)) {
57
#if CONFIG_HIGHBITDEPTH
Angie Chiang's avatar
Angie Chiang committed
58 59 60
    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
      aom_highbd_subtract_block_c(rows, cols, diff, diff_stride, src8,
                                  src_stride, pred8, pred_stride, xd->bd);
61 62
      return;
    }
63
#endif  // CONFIG_HIGHBITDEPTH
Angie Chiang's avatar
Angie Chiang committed
64 65
    aom_subtract_block_c(rows, cols, diff, diff_stride, src8, src_stride, pred8,
                         pred_stride);
66 67 68 69

    return;
  }

70
#if CONFIG_HIGHBITDEPTH
Angie Chiang's avatar
Angie Chiang committed
71 72 73
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
    aom_highbd_subtract_block(rows, cols, diff, diff_stride, src8, src_stride,
                              pred8, pred_stride, xd->bd);
Jingning Han's avatar
Jingning Han committed
74 75
    return;
  }
76
#endif  // CONFIG_HIGHBITDEPTH
Angie Chiang's avatar
Angie Chiang committed
77 78 79 80
  aom_subtract_block(rows, cols, diff, diff_stride, src8, src_stride, pred8,
                     pred_stride);
}

Angie Chiang's avatar
Angie Chiang committed
81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100
void av1_subtract_txb(MACROBLOCK *x, int plane, BLOCK_SIZE plane_bsize,
                      int blk_col, int blk_row, TX_SIZE tx_size) {
  MACROBLOCKD *const xd = &x->e_mbd;
  struct macroblock_plane *const p = &x->plane[plane];
  const struct macroblockd_plane *const pd = &x->e_mbd.plane[plane];
  const int diff_stride = block_size_wide[plane_bsize];
  const int src_stride = p->src.stride;
  const int dst_stride = pd->dst.stride;
  const int tx1d_width = tx_size_wide[tx_size];
  const int tx1d_height = tx_size_high[tx_size];
  uint8_t *dst =
      &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
  uint8_t *src =
      &p->src.buf[(blk_row * src_stride + blk_col) << tx_size_wide_log2[0]];
  int16_t *src_diff =
      &p->src_diff[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
  subtract_block(xd, tx1d_height, tx1d_width, src_diff, diff_stride, src,
                 src_stride, dst, dst_stride);
}

Angie Chiang's avatar
Angie Chiang committed
101 102 103 104 105 106 107 108 109 110
void av1_subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
  struct macroblock_plane *const p = &x->plane[plane];
  const struct macroblockd_plane *const pd = &x->e_mbd.plane[plane];
  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
  const int bw = block_size_wide[plane_bsize];
  const int bh = block_size_high[plane_bsize];
  const MACROBLOCKD *xd = &x->e_mbd;

  subtract_block(xd, bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
                 pd->dst.buf, pd->dst.stride);
Jingning Han's avatar
Jingning Han committed
111 112
}

113 114
// These numbers are empirically obtained.
static const int plane_rd_mult[REF_TYPES][PLANE_TYPES] = {
Thomas Davies's avatar
Thomas Davies committed
115
  { 10, 7 }, { 8, 5 },
116
};
Jingning Han's avatar
Jingning Han committed
117

118 119 120
static INLINE unsigned int get_token_bit_costs(
    unsigned int token_costs[2][COEFF_CONTEXTS][ENTROPY_TOKENS], int skip_eob,
    int ctx, int token) {
121 122 123 124
  (void)skip_eob;
  return token_costs[token == ZERO_TOKEN || token == EOB_TOKEN][ctx][token];
}

125
#if !CONFIG_LV_MAP
126

127 128
static int optimize_b_greedy(const AV1_COMMON *cm, MACROBLOCK *mb, int plane,
                             int block, TX_SIZE tx_size, int ctx) {
129 130 131 132 133 134 135 136 137 138 139 140
  MACROBLOCKD *const xd = &mb->e_mbd;
  struct macroblock_plane *const p = &mb->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
  const int ref = is_inter_block(&xd->mi[0]->mbmi);
  uint8_t token_cache[MAX_TX_SQUARE];
  const tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
  tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
  const int eob = p->eobs[block];
  const PLANE_TYPE plane_type = pd->plane_type;
  const int16_t *const dequant_ptr = pd->dequant;
  const uint8_t *const band_translate = get_band_translate(tx_size);
hui su's avatar
hui su committed
141
  const TX_TYPE tx_type = av1_get_tx_type(plane_type, xd, block, tx_size);
142
  const SCAN_ORDER *const scan_order =
Angie Chiang's avatar
Angie Chiang committed
143
      get_scan(cm, tx_size, tx_type, &xd->mi[0]->mbmi);
144 145 146
  const int16_t *const scan = scan_order->scan;
  const int16_t *const nb = scan_order->neighbors;
  int dqv;
147
  const int shift = av1_get_tx_scale(tx_size);
148 149
#if CONFIG_AOM_QM
  int seg_id = xd->mi[0]->mbmi.segment_id;
150 151 152 153 154
  // Use a flat matrix (i.e. no weighting) for 1D and Identity transforms
  const qm_val_t *iqmatrix =
      IS_2D_TRANSFORM(tx_type)
          ? pd->seg_iqmatrix[seg_id][!ref][tx_size]
          : cm->giqmatrix[NUM_QM_LEVELS - 1][0][0][tx_size];
155 156 157 158 159 160 161 162 163 164 165 166 167 168
#endif
#if CONFIG_NEW_QUANT
  int dq = get_dq_profile_from_ctx(mb->qindex, ctx, ref, plane_type);
  const dequant_val_type_nuq *dequant_val = pd->dequant_val_nuq[dq];
#endif  // CONFIG_NEW_QUANT
  int sz = 0;
  int64_t rd_cost0, rd_cost1;
  int16_t t0, t1;
  int i, final_eob;
  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, xd->bd);
  unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
      mb->token_costs[txsize_sqr_map[tx_size]][plane_type][ref];
  const int default_eob = tx_size_2d[tx_size];

169
  assert(mb->qindex > 0);
170 171 172 173 174 175 176 177 178

  assert((!plane_type && !plane) || (plane_type && plane));
  assert(eob <= default_eob);

  int64_t rdmult = (mb->rdmult * plane_rd_mult[ref][plane_type]) >> 1;

  int64_t rate0, rate1;
  for (i = 0; i < eob; i++) {
    const int rc = scan[i];
179
    token_cache[rc] = av1_pt_energy_class[av1_get_token(qcoeff[rc])];
180 181 182 183 184 185 186 187
  }

  unsigned int(*token_costs_ptr)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
      token_costs;

  final_eob = 0;

  int64_t eob_cost0, eob_cost1;
188 189
  tran_low_t before_best_eob_qc = 0;
  tran_low_t before_best_eob_dqc = 0;
190 191 192 193

  const int ctx0 = ctx;
  /* Record the r-d cost */
  int64_t accu_rate = 0;
194 195 196
  // Initialized to the worst possible error for the largest transform size.
  // This ensures that it never goes negative.
  int64_t accu_error = ((int64_t)1) << 50;
197 198 199

  rate0 = get_token_bit_costs(*(token_costs_ptr + band_translate[0]), 0, ctx0,
                              EOB_TOKEN);
200
  int64_t best_block_rd_cost = RDCOST(rdmult, rate0, accu_error);
201 202 203 204 205 206 207 208 209 210 211 212 213 214

  // int64_t best_block_rd_cost_all0 = best_block_rd_cost;
  int x_prev = 1;
  for (i = 0; i < eob; i++) {
    const int rc = scan[i];
    int x = qcoeff[rc];
    sz = -(x < 0);

    int band_cur = band_translate[i];
    int ctx_cur = (i == 0) ? ctx : get_coef_context(nb, token_cache, i);
    int token_tree_sel_cur = (x_prev == 0);

    if (x == 0) {
      // no need to search when x == 0
215 216 217
      int token = av1_get_token(x);
      rate0 = get_token_bit_costs(*(token_costs_ptr + band_cur),
                                  token_tree_sel_cur, ctx_cur, token);
218 219 220 221 222 223 224 225
      accu_rate += rate0;
      x_prev = 0;
      // accu_error does not change when x==0
    } else {
      /*  Computing distortion
       */
      // compute the distortion for the first candidate
      // and the distortion for quantizing to 0.
226
      int dx0 = abs(coeff[rc]) * (1 << shift);
227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249
#if CONFIG_HIGHBITDEPTH
      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
        dx0 >>= xd->bd - 8;
      }
#endif
      int64_t d0 = (int64_t)dx0 * dx0;

      int x_a = x - 2 * sz - 1;
      int64_t d2, d2_a;

      int dx;

#if CONFIG_AOM_QM
      int iwt = iqmatrix[rc];
      dqv = dequant_ptr[rc != 0];
      dqv = ((iwt * (int)dqv) + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
#else
      dqv = dequant_ptr[rc != 0];
#endif

      dx = (dqcoeff[rc] - coeff[rc]) * (1 << shift);
#if CONFIG_HIGHBITDEPTH
      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
250 251 252
        int dx_sign = dx < 0 ? 1 : 0;
        dx = abs(dx) >> (xd->bd - 8);
        if (dx_sign) dx = -dx;
253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307
      }
#endif  // CONFIG_HIGHBITDEPTH
      d2 = (int64_t)dx * dx;

      /* compute the distortion for the second candidate
       * x_a = x - 2 * sz + 1;
       */
      if (x_a != 0) {
#if CONFIG_NEW_QUANT
        dx = av1_dequant_coeff_nuq(x, dqv, dequant_val[band_translate[i]]) -
             (coeff[rc] << shift);
#if CONFIG_HIGHBITDEPTH
        if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
          dx >>= xd->bd - 8;
        }
#endif  // CONFIG_HIGHBITDEPTH
#else   // CONFIG_NEW_QUANT
#if CONFIG_HIGHBITDEPTH
        if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
          dx -= ((dqv >> (xd->bd - 8)) + sz) ^ sz;
        } else {
          dx -= (dqv + sz) ^ sz;
        }
#else
        dx -= (dqv + sz) ^ sz;
#endif  // CONFIG_HIGHBITDEPTH
#endif  // CONFIG_NEW_QUANT
        d2_a = (int64_t)dx * dx;
      } else {
        d2_a = d0;
      }
      /*  Computing rates and r-d cost
       */

      int best_x, best_eob_x;
      int64_t base_bits, next_bits0, next_bits1;
      int64_t next_eob_bits0, next_eob_bits1;

      // rate cost of x
      base_bits = av1_get_token_cost(x, &t0, cat6_bits);
      rate0 = base_bits + get_token_bit_costs(*(token_costs_ptr + band_cur),
                                              token_tree_sel_cur, ctx_cur, t0);

      base_bits = av1_get_token_cost(x_a, &t1, cat6_bits);
      rate1 = base_bits + get_token_bit_costs(*(token_costs_ptr + band_cur),
                                              token_tree_sel_cur, ctx_cur, t1);

      next_bits0 = 0;
      next_bits1 = 0;
      next_eob_bits0 = 0;
      next_eob_bits1 = 0;

      if (i < default_eob - 1) {
        int ctx_next, token_tree_sel_next;
        int band_next = band_translate[i + 1];
308 309
        int token_next =
            i + 1 != eob ? av1_get_token(qcoeff[scan[i + 1]]) : EOB_TOKEN;
310 311 312 313 314

        token_cache[rc] = av1_pt_energy_class[t0];
        ctx_next = get_coef_context(nb, token_cache, i + 1);
        token_tree_sel_next = (x == 0);

315 316 317
        next_bits0 =
            get_token_bit_costs(*(token_costs_ptr + band_next),
                                token_tree_sel_next, ctx_next, token_next);
318 319 320 321 322 323 324 325
        next_eob_bits0 =
            get_token_bit_costs(*(token_costs_ptr + band_next),
                                token_tree_sel_next, ctx_next, EOB_TOKEN);

        token_cache[rc] = av1_pt_energy_class[t1];
        ctx_next = get_coef_context(nb, token_cache, i + 1);
        token_tree_sel_next = (x_a == 0);

326 327 328
        next_bits1 =
            get_token_bit_costs(*(token_costs_ptr + band_next),
                                token_tree_sel_next, ctx_next, token_next);
329 330 331 332 333 334 335 336

        if (x_a != 0) {
          next_eob_bits1 =
              get_token_bit_costs(*(token_costs_ptr + band_next),
                                  token_tree_sel_next, ctx_next, EOB_TOKEN);
        }
      }

337 338
      rd_cost0 = RDCOST(rdmult, (rate0 + next_bits0), d2);
      rd_cost1 = RDCOST(rdmult, (rate1 + next_bits1), d2_a);
339 340 341

      best_x = (rd_cost1 < rd_cost0);

342
      eob_cost0 = RDCOST(rdmult, (accu_rate + rate0 + next_eob_bits0),
343 344 345
                         (accu_error + d2 - d0));
      eob_cost1 = eob_cost0;
      if (x_a != 0) {
346
        eob_cost1 = RDCOST(rdmult, (accu_rate + rate1 + next_eob_bits1),
347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363
                           (accu_error + d2_a - d0));
        best_eob_x = (eob_cost1 < eob_cost0);
      } else {
        best_eob_x = 0;
      }

      int dqc, dqc_a = 0;

      dqc = dqcoeff[rc];
      if (best_x + best_eob_x) {
        if (x_a != 0) {
#if CONFIG_NEW_QUANT
          dqc_a = av1_dequant_abscoeff_nuq(abs(x_a), dqv,
                                           dequant_val[band_translate[i]]);
          dqc_a = shift ? ROUND_POWER_OF_TWO(dqc_a, shift) : dqc_a;
          if (sz) dqc_a = -dqc_a;
#else
364 365
          if (x_a < 0)
            dqc_a = -((-x_a * dqv) >> shift);
366
          else
367
            dqc_a = (x_a * dqv) >> shift;
368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390
#endif  // CONFIG_NEW_QUANT
        } else {
          dqc_a = 0;
        }  // if (x_a != 0)
      }

      // record the better quantized value
      if (best_x) {
        qcoeff[rc] = x_a;
        dqcoeff[rc] = dqc_a;

        accu_rate += rate1;
        accu_error += d2_a - d0;
        assert(d2_a <= d0);

        token_cache[rc] = av1_pt_energy_class[t1];
      } else {
        accu_rate += rate0;
        accu_error += d2 - d0;
        assert(d2 <= d0);

        token_cache[rc] = av1_pt_energy_class[t0];
      }
391
      assert(accu_error >= 0);
392 393 394 395

      x_prev = qcoeff[rc];

      // determine whether to move the eob position to i+1
396 397
      int use_a = (x_a != 0) && (best_eob_x);
      int64_t best_eob_cost_i = use_a ? eob_cost1 : eob_cost0;
398 399 400 401

      if (best_eob_cost_i < best_block_rd_cost) {
        best_block_rd_cost = best_eob_cost_i;
        final_eob = i + 1;
402 403 404 405 406 407 408
        if (use_a) {
          before_best_eob_qc = x_a;
          before_best_eob_dqc = dqc_a;
        } else {
          before_best_eob_qc = x;
          before_best_eob_dqc = dqc;
        }
409 410 411 412 413 414
      }
    }  // if (x==0)
  }    // for (i)

  assert(final_eob <= eob);
  if (final_eob > 0) {
415
    assert(before_best_eob_qc != 0);
416 417
    i = final_eob - 1;
    int rc = scan[i];
418 419
    qcoeff[rc] = before_best_eob_qc;
    dqcoeff[rc] = before_best_eob_dqc;
420 421 422 423 424 425 426 427 428 429 430
  }

  for (i = final_eob; i < eob; i++) {
    int rc = scan[i];
    qcoeff[rc] = 0;
    dqcoeff[rc] = 0;
  }

  mb->plane[plane].eobs[block] = final_eob;
  return final_eob;
}
431
#endif  // !CONFIG_LV_MAP
432

Jingning Han's avatar
Jingning Han committed
433 434 435 436
int av1_optimize_b(const AV1_COMMON *cm, MACROBLOCK *mb, int plane, int blk_row,
                   int blk_col, int block, BLOCK_SIZE plane_bsize,
                   TX_SIZE tx_size, const ENTROPY_CONTEXT *a,
                   const ENTROPY_CONTEXT *l) {
437 438 439 440 441 442
  MACROBLOCKD *const xd = &mb->e_mbd;
  struct macroblock_plane *const p = &mb->plane[plane];
  const int eob = p->eobs[block];
  assert((mb->qindex == 0) ^ (xd->lossless[xd->mi[0]->mbmi.segment_id] == 0));
  if (eob == 0) return eob;
  if (xd->lossless[xd->mi[0]->mbmi.segment_id]) return eob;
Jingning Han's avatar
Jingning Han committed
443

444
#if CONFIG_PVQ
445 446
  (void)cm;
  (void)tx_size;
447 448
  (void)a;
  (void)l;
449 450
  return eob;
#endif
Jingning Han's avatar
Jingning Han committed
451

452
#if !CONFIG_LV_MAP
453
  (void)plane_bsize;
Jingning Han's avatar
Jingning Han committed
454 455
  (void)blk_row;
  (void)blk_col;
456 457 458 459
#if CONFIG_VAR_TX
  int ctx = get_entropy_context(tx_size, a, l);
#else
  int ctx = combine_entropy_contexts(*a, *l);
460
#endif  // CONFIG_VAR_TX
461
  return optimize_b_greedy(cm, mb, plane, block, tx_size, ctx);
462 463 464
#else   // !CONFIG_LV_MAP
  TXB_CTX txb_ctx;
  get_txb_ctx(plane_bsize, tx_size, plane, a, l, &txb_ctx);
Jingning Han's avatar
Jingning Han committed
465 466
  return av1_optimize_txb(cm, mb, plane, blk_row, blk_col, block, tx_size,
                          &txb_ctx);
467
#endif  // !CONFIG_LV_MAP
468
}
469

Thomas Daede's avatar
Thomas Daede committed
470
#if !CONFIG_PVQ
Angie Chiang's avatar
Angie Chiang committed
471 472
typedef enum QUANT_FUNC {
  QUANT_FUNC_LOWBD = 0,
473
  QUANT_FUNC_HIGHBD = 1,
474
  QUANT_FUNC_TYPES = 2
Angie Chiang's avatar
Angie Chiang committed
475 476
} QUANT_FUNC;

477 478
static AV1_QUANT_FACADE
    quant_func_list[AV1_XFORM_QUANT_TYPES][QUANT_FUNC_TYPES] = {
479
#if !CONFIG_NEW_QUANT
480
      { av1_quantize_fp_facade, av1_highbd_quantize_fp_facade },
Yaowu Xu's avatar
Yaowu Xu committed
481 482
      { av1_quantize_b_facade, av1_highbd_quantize_b_facade },
      { av1_quantize_dc_facade, av1_highbd_quantize_dc_facade },
483
#else   // !CONFIG_NEW_QUANT
484 485 486
      { av1_quantize_fp_nuq_facade, av1_highbd_quantize_fp_nuq_facade },
      { av1_quantize_b_nuq_facade, av1_highbd_quantize_b_nuq_facade },
      { av1_quantize_dc_nuq_facade, av1_highbd_quantize_dc_nuq_facade },
487
#endif  // !CONFIG_NEW_QUANT
488 489
      { NULL, NULL }
    };
490
#endif  // !CONFIG_PVQ
491

492 493 494 495
typedef void (*fwdTxfmFunc)(const int16_t *diff, tran_low_t *coeff, int stride,
                            FWD_TXFM_PARAM *param);
static const fwdTxfmFunc fwd_txfm_func[2] = { av1_fwd_txfm,
                                              av1_highbd_fwd_txfm };
496

Angie Chiang's avatar
Angie Chiang committed
497 498
void av1_xform_quant(const AV1_COMMON *cm, MACROBLOCK *x, int plane, int block,
                     int blk_row, int blk_col, BLOCK_SIZE plane_bsize,
499 500
                     TX_SIZE tx_size, int ctx,
                     AV1_XFORM_QUANT xform_quant_idx) {
Jingning Han's avatar
Jingning Han committed
501
  MACROBLOCKD *const xd = &x->e_mbd;
502
  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
Yushin Cho's avatar
Yushin Cho committed
503
#if !(CONFIG_PVQ || CONFIG_DAALA_DIST)
Jingning Han's avatar
Jingning Han committed
504 505
  const struct macroblock_plane *const p = &x->plane[plane];
  const struct macroblockd_plane *const pd = &xd->plane[plane];
506 507 508 509
#else
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
#endif
510
  PLANE_TYPE plane_type = get_plane_type(plane);
hui su's avatar
hui su committed
511
  const TX_TYPE tx_type = av1_get_tx_type(plane_type, xd, block, tx_size);
Angie Chiang's avatar
Angie Chiang committed
512 513

#if CONFIG_AOM_QM || CONFIG_NEW_QUANT
514
  const int is_inter = is_inter_block(mbmi);
Angie Chiang's avatar
Angie Chiang committed
515 516 517
#endif

  const SCAN_ORDER *const scan_order = get_scan(cm, tx_size, tx_type, mbmi);
Jingning Han's avatar
Jingning Han committed
518 519 520 521
  tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
  tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
  uint16_t *const eob = &p->eobs[block];
522
  const int diff_stride = block_size_wide[plane_bsize];
523
#if CONFIG_AOM_QM
524
  int seg_id = mbmi->segment_id;
525 526 527 528 529 530 531 532
  // Use a flat matrix (i.e. no weighting) for 1D and Identity transforms
  const qm_val_t *qmatrix =
      IS_2D_TRANSFORM(tx_type) ? pd->seg_qmatrix[seg_id][!is_inter][tx_size]
                               : cm->gqmatrix[NUM_QM_LEVELS - 1][0][0][tx_size];
  const qm_val_t *iqmatrix =
      IS_2D_TRANSFORM(tx_type)
          ? pd->seg_iqmatrix[seg_id][!is_inter][tx_size]
          : cm->giqmatrix[NUM_QM_LEVELS - 1][0][0][tx_size];
533
#endif
Angie Chiang's avatar
Angie Chiang committed
534 535

  FWD_TXFM_PARAM fwd_txfm_param;
536

Lester Lu's avatar
Lester Lu committed
537
#if CONFIG_PVQ || CONFIG_DAALA_DIST || CONFIG_LGT
Yushin Cho's avatar
Yushin Cho committed
538 539
  uint8_t *dst;
  const int dst_stride = pd->dst.stride;
Lester Lu's avatar
Lester Lu committed
540 541
#if CONFIG_PVQ || CONFIG_DAALA_DIST
  int16_t *pred;
542 543
  const int txw = tx_size_wide[tx_size];
  const int txh = tx_size_high[tx_size];
Yushin Cho's avatar
Yushin Cho committed
544 545
  int i, j;
#endif
Lester Lu's avatar
Lester Lu committed
546
#endif
Yushin Cho's avatar
Yushin Cho committed
547

548 549
#if !CONFIG_PVQ
  const int tx2d_size = tx_size_2d[tx_size];
550
  QUANT_PARAM qparam;
551 552
  const int16_t *src_diff;

553 554
  src_diff =
      &p->src_diff[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
555
  qparam.log_scale = av1_get_tx_scale(tx_size);
556 557 558 559 560 561 562 563
#if CONFIG_NEW_QUANT
  qparam.tx_size = tx_size;
  qparam.dq = get_dq_profile_from_ctx(x->qindex, ctx, is_inter, plane_type);
#endif  // CONFIG_NEW_QUANT
#if CONFIG_AOM_QM
  qparam.qmatrix = qmatrix;
  qparam.iqmatrix = iqmatrix;
#endif  // CONFIG_AOM_QM
564 565 566 567
#else
  tran_low_t *ref_coeff = BLOCK_OFFSET(pd->pvq_ref_coeff, block);
  int skip = 1;
  PVQ_INFO *pvq_info = NULL;
Yushin Cho's avatar
Yushin Cho committed
568 569 570
  uint8_t *src;
  int16_t *src_int16;
  const int src_stride = p->src.stride;
571

572
  (void)ctx;
573 574 575 576 577 578 579
  (void)scan_order;
  (void)qcoeff;

  if (x->pvq_coded) {
    assert(block < MAX_PVQ_BLOCKS_IN_SB);
    pvq_info = &x->pvq[block][plane];
  }
580 581 582
  src = &p->src.buf[(blk_row * src_stride + blk_col) << tx_size_wide_log2[0]];
  src_int16 =
      &p->src_int16[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
Yushin Cho's avatar
Yushin Cho committed
583

584
#if CONFIG_HIGHBITDEPTH
Thomas Daede's avatar
Thomas Daede committed
585
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
586 587
    for (j = 0; j < txh; j++)
      for (i = 0; i < txw; i++)
Thomas Daede's avatar
Thomas Daede committed
588 589 590
        src_int16[diff_stride * j + i] =
            CONVERT_TO_SHORTPTR(src)[src_stride * j + i];
  } else {
591
#endif  // CONFIG_HIGHBITDEPTH
592 593
    for (j = 0; j < txh; j++)
      for (i = 0; i < txw; i++)
Thomas Daede's avatar
Thomas Daede committed
594
        src_int16[diff_stride * j + i] = src[src_stride * j + i];
595
#if CONFIG_HIGHBITDEPTH
Thomas Daede's avatar
Thomas Daede committed
596
  }
597
#endif  // CONFIG_HIGHBITDEPTH
Yushin Cho's avatar
Yushin Cho committed
598 599
#endif

Lester Lu's avatar
Lester Lu committed
600
#if CONFIG_PVQ || CONFIG_DAALA_DIST || CONFIG_LGT
Yushin Cho's avatar
Yushin Cho committed
601
  dst = &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
Lester Lu's avatar
Lester Lu committed
602
#if CONFIG_PVQ || CONFIG_DAALA_DIST
603
  pred = &pd->pred[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
604

Thomas Daede's avatar
Thomas Daede committed
605 606
// copy uint8 orig and predicted block to int16 buffer
// in order to use existing VP10 transform functions
607
#if CONFIG_HIGHBITDEPTH
Thomas Daede's avatar
Thomas Daede committed
608
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
609 610
    for (j = 0; j < txh; j++)
      for (i = 0; i < txw; i++)
Thomas Daede's avatar
Thomas Daede committed
611 612 613
        pred[diff_stride * j + i] =
            CONVERT_TO_SHORTPTR(dst)[dst_stride * j + i];
  } else {
614
#endif  // CONFIG_HIGHBITDEPTH
615 616
    for (j = 0; j < txh; j++)
      for (i = 0; i < txw; i++)
Thomas Daede's avatar
Thomas Daede committed
617
        pred[diff_stride * j + i] = dst[dst_stride * j + i];
618
#if CONFIG_HIGHBITDEPTH
Thomas Daede's avatar
Thomas Daede committed
619
  }
620
#endif  // CONFIG_HIGHBITDEPTH
Lester Lu's avatar
Lester Lu committed
621 622
#endif  // CONFIG_PVQ || CONFIG_DAALA_DIST
#endif  // CONFIG_PVQ || CONFIG_DAALA_DIST || CONFIG_LGT
Yushin Cho's avatar
Yushin Cho committed
623

624
  (void)ctx;
625 626

  fwd_txfm_param.tx_type = tx_type;
Angie Chiang's avatar
Angie Chiang committed
627
  fwd_txfm_param.tx_size = tx_size;
628
  fwd_txfm_param.lossless = xd->lossless[mbmi->segment_id];
Lester Lu's avatar
Lester Lu committed
629 630 631 632 633 634
#if CONFIG_LGT
  fwd_txfm_param.is_inter = is_inter_block(mbmi);
  fwd_txfm_param.dst = dst;
  fwd_txfm_param.stride = dst_stride;
  fwd_txfm_param.mode = get_prediction_mode(xd->mi[0], plane, tx_size, block);
#endif
Angie Chiang's avatar
Angie Chiang committed
635

Thomas Daede's avatar
Thomas Daede committed
636
#if !CONFIG_PVQ
637
  fwd_txfm_param.bd = xd->bd;
638 639 640
  const int is_hbd = get_bitdepth_data_path_index(xd);
  fwd_txfm_func[is_hbd](src_diff, coeff, diff_stride, &fwd_txfm_param);

Yaowu Xu's avatar
Yaowu Xu committed
641
  if (xform_quant_idx != AV1_XFORM_QUANT_SKIP_QUANT) {
642
    if (LIKELY(!x->skip_block)) {
643
      quant_func_list[xform_quant_idx][is_hbd](
644
          coeff, tx2d_size, p, qcoeff, pd, dqcoeff, eob, scan_order, &qparam);
645
    } else {
Yaowu Xu's avatar
Yaowu Xu committed
646
      av1_quantize_skip(tx2d_size, qcoeff, dqcoeff, eob);
Jingning Han's avatar
Jingning Han committed
647 648
    }
  }
649 650 651 652
#if CONFIG_LV_MAP
  p->txb_entropy_ctx[block] =
      (uint8_t)av1_get_txb_entropy_context(qcoeff, scan_order, *eob);
#endif  // CONFIG_LV_MAP
653 654
  return;
#else  // CONFIG_PVQ
Angie Chiang's avatar
Angie Chiang committed
655
  (void)xform_quant_idx;
656
#if CONFIG_HIGHBITDEPTH
Thomas Daede's avatar
Thomas Daede committed
657 658
  fwd_txfm_param.bd = xd->bd;
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
hui su's avatar
hui su committed
659 660
    av1_highbd_fwd_txfm(src_int16, coeff, diff_stride, &fwd_txfm_param);
    av1_highbd_fwd_txfm(pred, ref_coeff, diff_stride, &fwd_txfm_param);
Thomas Daede's avatar
Thomas Daede committed
661 662
  } else {
#endif
hui su's avatar
hui su committed
663 664
    av1_fwd_txfm(src_int16, coeff, diff_stride, &fwd_txfm_param);
    av1_fwd_txfm(pred, ref_coeff, diff_stride, &fwd_txfm_param);
665
#if CONFIG_HIGHBITDEPTH
Thomas Daede's avatar
Thomas Daede committed
666 667
  }
#endif
668 669

  // PVQ for inter mode block
670
  if (!x->skip_block) {
ltrudeau's avatar
ltrudeau committed
671
    PVQ_SKIP_TYPE ac_dc_coded =
Thomas Daede's avatar
Thomas Daede committed
672
        av1_pvq_encode_helper(x,
ltrudeau's avatar
ltrudeau committed
673 674 675 676 677 678 679 680 681 682 683 684
                              coeff,        // target original vector
                              ref_coeff,    // reference vector
                              dqcoeff,      // de-quantized vector
                              eob,          // End of Block marker
                              pd->dequant,  // aom's quantizers
                              plane,        // image plane
                              tx_size,      // block size in log_2 - 2
                              tx_type,
                              &x->rate,  // rate measured
                              x->pvq_speed,
                              pvq_info);  // PVQ info for a block
    skip = ac_dc_coded == PVQ_SKIP;
685
  }
686 687 688 689
  x->pvq_skip[plane] = skip;

  if (!skip) mbmi->skip = 0;
#endif  // #if !CONFIG_PVQ
Jingning Han's avatar
Jingning Han committed
690 691
}

692
static void encode_block(int plane, int block, int blk_row, int blk_col,
693
                         BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
Jingning Han's avatar
Jingning Han committed
694
  struct encode_b_args *const args = arg;
Angie Chiang's avatar
Angie Chiang committed
695
  AV1_COMMON *cm = args->cm;
Jingning Han's avatar
Jingning Han committed
696 697
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
698
  int ctx;
Jingning Han's avatar
Jingning Han committed
699 700 701 702
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
  uint8_t *dst;
703
#if !CONFIG_PVQ
Jingning Han's avatar
Jingning Han committed
704
  ENTROPY_CONTEXT *a, *l;
705
#endif
706
#if CONFIG_VAR_TX
707
  int bw = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
708
#endif
709 710
  dst = &pd->dst
             .buf[(blk_row * pd->dst.stride + blk_col) << tx_size_wide_log2[0]];
711 712

#if !CONFIG_PVQ
713 714
  a = &args->ta[blk_col];
  l = &args->tl[blk_row];
715 716 717 718 719
#if CONFIG_VAR_TX
  ctx = get_entropy_context(tx_size, a, l);
#else
  ctx = combine_entropy_contexts(*a, *l);
#endif
720 721 722
#else
  ctx = 0;
#endif  // CONFIG_PVQ
Jingning Han's avatar
Jingning Han committed
723

Alex Converse's avatar
Alex Converse committed
724
#if CONFIG_VAR_TX
725
  // Assert not magic number (uninitialized).
726
  assert(x->blk_skip[plane][blk_row * bw + blk_col] != 234);
727

728
  if (x->blk_skip[plane][blk_row * bw + blk_col] == 0) {
729
#else
730
  {
731
#endif
Angie Chiang's avatar
Angie Chiang committed
732
    av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
733
                    ctx, AV1_XFORM_QUANT_FP);
Jingning Han's avatar
Jingning Han committed
734
  }
735 736
#if CONFIG_VAR_TX
  else {
737
    p->eobs[block] = 0;
738 739
  }
#endif
740

741
#if !CONFIG_PVQ
Jingning Han's avatar
Jingning Han committed
742 743
  av1_optimize_b(cm, x, plane, blk_row, blk_col, block, plane_bsize, tx_size, a,
                 l);
744

745 746
  av1_set_txb_context(x, plane, block, tx_size, a, l);

747
  if (p->eobs[block]) *(args->skip) = 0;
Jingning Han's avatar
Jingning Han committed
748

749
  if (p->eobs[block] == 0) return;
750 751 752 753 754 755
#else
  (void)ctx;
  if (!x->pvq_skip[plane]) *(args->skip) = 0;

  if (x->pvq_skip[plane]) return;
#endif
hui su's avatar
hui su committed
756
  const TX_TYPE tx_type = av1_get_tx_type(pd->plane_type, xd, block, tx_size);
Lester Lu's avatar
Lester Lu committed
757 758 759 760 761
#if CONFIG_LGT
  PREDICTION_MODE mode = get_prediction_mode(xd->mi[0], plane, tx_size, block);
  av1_inverse_transform_block(xd, dqcoeff, mode, tx_type, tx_size, dst,
                              pd->dst.stride, p->eobs[block]);
#else
762 763
  av1_inverse_transform_block(xd, dqcoeff, tx_type, tx_size, dst,
                              pd->dst.stride, p->eobs[block]);
Lester Lu's avatar
Lester Lu committed
764
#endif
Jingning Han's avatar
Jingning Han committed
765 766
}

767 768 769 770 771 772 773 774
#if CONFIG_VAR_TX
static void encode_block_inter(int plane, int block, int blk_row, int blk_col,
                               BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
                               void *arg) {
  struct encode_b_args *const args = arg;
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
775
  const BLOCK_SIZE bsize = txsize_to_bsize[tx_size];
776
  const struct macroblockd_plane *const pd = &xd->plane[plane];
777 778
  const int tx_row = blk_row >> (1 - pd->subsampling_y);
  const int tx_col = blk_col >> (1 - pd->subsampling_x);
Debargha Mukherjee's avatar
Debargha Mukherjee committed
779
  TX_SIZE plane_tx_size;
780 781
  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
782

783
  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
784

785 786 787
  plane_tx_size =
      plane ? uv_txsize_lookup[bsize][mbmi->inter_tx_size[tx_row][tx_col]][0][0]
            : mbmi->inter_tx_size[tx_row][tx_col];
Debargha Mukherjee's avatar
Debargha Mukherjee committed
788

789
  if (tx_size == plane_tx_size) {
790
    encode_block(plane, block, blk_row, blk_col, plane_bsize, tx_size, arg);
791
  } else {
792
    assert(tx_size < TX_SIZES_ALL);
793
    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
794
    assert(sub_txs < tx_size);
795 796
    // This is the square transform block partition entry point.
    int bsl = tx_size_wide_unit[sub_txs];
797 798
    int i;
    assert(bsl > 0);
799

800
    for (i = 0; i < 4; ++i) {
801 802 803
      const int offsetr = blk_row + ((i >> 1) * bsl);
      const int offsetc = blk_col + ((i & 0x01) * bsl);
      int step = tx_size_wide_unit[sub_txs] * tx_size_high_unit[sub_txs];
804

805
      if (offsetr >= max_blocks_high || offsetc >= max_blocks_wide) continue;
806

807 808 809
      encode_block_inter(plane, block, offsetr, offsetc, plane_bsize, sub_txs,
                         arg);
      block += step;
810 811 812 813 814
    }
  }
}
#endif

Angie Chiang's avatar
Angie Chiang committed
815 816 817 818 819
typedef struct encode_block_pass1_args {
  AV1_COMMON *cm;
  MACROBLOCK *x;
} encode_block_pass1_args;

820
static void encode_block_pass1(int plane, int block, int blk_row, int blk_col,
821 822
                               BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
                               void *arg) {
Angie Chiang's avatar
Angie Chiang committed
823 824 825
  encode_block_pass1_args *args = (encode_block_pass1_args *)arg;
  AV1_COMMON *cm = args->cm;
  MACROBLOCK *const x = args->x;
Jingning Han's avatar
Jingning Han committed
826 827 828 829
  MACROBLOCKD *const xd = &x->e_mbd;
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
830
  INV_TXFM_PARAM inv_txfm_param;
Jingning Han's avatar
Jingning Han committed
831
  uint8_t *dst;
832
  int ctx = 0;
833 834
  dst = &pd->dst
             .buf[(blk_row * pd->dst.stride + blk_col) << tx_size_wide_log2[0]];
Jingning Han's avatar
Jingning Han committed
835

Angie Chiang's avatar
Angie Chiang committed
836
  av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
837
                  ctx, AV1_XFORM_QUANT_B);
838
#if !CONFIG_PVQ
Jingning Han's avatar
Jingning Han committed
839
  if (p->eobs[block] > 0) {
840 841 842 843 844 845 846 847
#else
  if (!x->pvq_skip[plane]) {
    {
      int tx_blk_size;
      int i, j;
      // transform block size in pixels
      tx_blk_size = tx_size_wide[tx_size];

Thomas Daede's avatar
Thomas Daede committed
848 849 850 851
// Since av1 does not have separate function which does inverse transform
// but av1_inv_txfm_add_*x*() also does addition of predicted image to
// inverse transformed image,
// pass blank dummy image to av1_inv_txfm_add_*x*(), i.e. set dst as zeros
852
#if CONFIG_HIGHBITDEPTH
Thomas Daede's avatar
Thomas Daede committed
853 854 855 856 857
      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
        for (j = 0; j < tx_blk_size; j++)
          for (i = 0; i < tx_blk_size; i++)
            CONVERT_TO_SHORTPTR(dst)[j * pd->dst.stride + i] = 0;
      } else {
858
#endif  // CONFIG_HIGHBITDEPTH
Thomas Daede's avatar
Thomas Daede committed
859 860
        for (j = 0; j < tx_blk_size; j++)
          for (i = 0; i < tx_blk_size; i++) dst[j * pd->dst.stride + i] = 0;
861
#if CONFIG_HIGHBITDEPTH
Thomas Daede's avatar
Thomas Daede committed
862
      }
863
#endif  // CONFIG_HIGHBITDEPTH
864
    }
Jingning Han's avatar
Jingning Han committed
865
#endif  // !CONFIG_PVQ
866 867 868 869
    inv_txfm_param.bd = xd->bd;
    inv_txfm_param.tx_type = DCT_DCT;
    inv_txfm_param.eob = p->eobs[block];
    inv_txfm_param.lossless = xd->lossless[xd->mi[0]->mbmi.segment_id];
870
#if CONFIG_HIGHBITDEPTH
Jingning Han's avatar
Jingning Han committed
871
    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
872 873
      av1_highbd_inv_txfm_add_4x4(dqcoeff, dst, pd->dst.stride,
                                  &inv_txfm_param);
874
      return;
Jingning Han's avatar
Jingning Han committed
875
    }
876
#endif  //  CONFIG_HIGHBITDEPTH
Yaowu Xu's avatar
Yaowu Xu committed
877
    if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {
878
      av1_iwht4x4_add(dqcoeff, dst, pd->dst.stride, &inv_txfm_param);
879
    } else {
880
      av1_idct4x4_add(dqcoeff, dst, pd->dst.stride, &inv_txfm_param);
881
    }
Jingning Han's avatar
Jingning Han committed
882 883 884
  }
}

Angie Chiang's avatar
Angie Chiang committed
885 886
void av1_encode_sby_pass1(AV1_COMMON *cm, MACROBLOCK *x, BLOCK_SIZE bsize) {
  encode_block_pass1_args args = { cm, x };
Yaowu Xu's avatar
Yaowu Xu committed
887 888
  av1_subtract_plane(x, bsize, 0);
  av1_foreach_transformed_block_in_plane(&x->e_mbd, bsize, 0,
Angie Chiang's avatar
Angie Chiang committed
889
                                         encode_block_pass1, &args);
Jingning Han's avatar
Jingning Han committed
890 891
}

Yaowu Xu's avatar
Yaowu Xu committed
892 893
void av1_encode_sb(AV1_COMMON *cm, MACROBLOCK *x, BLOCK_SIZE bsize, int mi_row,
                   int mi_col) {
Jingning Han's avatar
Jingning Han committed
894 895 896
  MACROBLOCKD *const xd = &x