encodemb.c 34.7 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
#include "av1/encoder/hybrid_fwd_txfm.h"
32 33 34
#if CONFIG_DAALA_TX
#include "av1/common/daala_inv_txfm.h"
#endif
35 36
#include "av1/encoder/rd.h"
#include "av1/encoder/tokenize.h"
Jingning Han's avatar
Jingning Han committed
37

38 39 40 41
#if CONFIG_CFL
#include "av1/common/cfl.h"
#endif

42 43 44
// 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
45 46 47 48
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) {
49
#if !CONFIG_HIGHBITDEPTH
Angie Chiang's avatar
Angie Chiang committed
50 51
  (void)xd;
#endif
Jingning Han's avatar
Jingning Han committed
52

Angie Chiang's avatar
Angie Chiang committed
53
  if (check_subtract_block_size(rows, cols)) {
54
#if CONFIG_HIGHBITDEPTH
Angie Chiang's avatar
Angie Chiang committed
55 56 57
    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);
58 59
      return;
    }
60
#endif  // CONFIG_HIGHBITDEPTH
Angie Chiang's avatar
Angie Chiang committed
61 62
    aom_subtract_block_c(rows, cols, diff, diff_stride, src8, src_stride, pred8,
                         pred_stride);
63 64 65 66

    return;
  }

67
#if CONFIG_HIGHBITDEPTH
Angie Chiang's avatar
Angie Chiang committed
68 69 70
  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
71 72
    return;
  }
73
#endif  // CONFIG_HIGHBITDEPTH
Angie Chiang's avatar
Angie Chiang committed
74 75 76 77
  aom_subtract_block(rows, cols, diff, diff_stride, src8, src_stride, pred8,
                     pred_stride);
}

Angie Chiang's avatar
Angie Chiang committed
78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97
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
98 99 100 101 102 103 104 105 106 107
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
108 109
}

110
#if !CONFIG_LV_MAP
111 112 113 114 115 116 117 118 119 120 121
// Shifting negative values is undefined behaviour in C99,
// and could mislead the optimizer, who might assume the shifted is positive.
// This also avoids ubsan warnings.
// In practise, this gets inlined by the optimizer to a single instruction.
static INLINE int signed_shift_right(int x, int shift) {
  if (x >= 0)
    return x >> shift;
  else
    return -((-x) >> shift);
}

122 123
// These numbers are empirically obtained.
static const int plane_rd_mult[REF_TYPES][PLANE_TYPES] = {
Thomas Davies's avatar
Thomas Davies committed
124
  { 10, 7 }, { 8, 5 },
125
};
Jingning Han's avatar
Jingning Han committed
126

127
static int optimize_b_greedy(const AV1_COMMON *cm, MACROBLOCK *mb, int plane,
128
                             int blk_row, int blk_col, int block,
129
                             TX_SIZE tx_size, int ctx, int fast_mode) {
130 131 132
  MACROBLOCKD *const xd = &mb->e_mbd;
  struct macroblock_plane *const p = &mb->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
133 134 135 136 137
  const PLANE_TYPE plane_type = pd->plane_type;
  const int eob = p->eobs[block];
  assert(mb->qindex > 0);
  assert((!plane_type && !plane) || (plane_type && plane));
  assert(eob <= tx_size_2d[tx_size]);
138 139 140 141
  const int ref = is_inter_block(&xd->mi[0]->mbmi);
  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);
142
  const int16_t *const dequant_ptr = p->dequant_QTX;
143
  const uint8_t *const band_translate = get_band_translate(tx_size);
144
  const TX_TYPE tx_type =
145
      av1_get_tx_type(plane_type, xd, blk_row, blk_col, block, tx_size);
146
  const SCAN_ORDER *const scan_order =
Angie Chiang's avatar
Angie Chiang committed
147
      get_scan(cm, tx_size, tx_type, &xd->mi[0]->mbmi);
148 149
  const int16_t *const scan = scan_order->scan;
  const int16_t *const nb = scan_order->neighbors;
150 151 152 153 154 155
#if CONFIG_DAALA_TX
  // This is one of the few places where RDO is done on coeffs; it
  // expects the coeffs to be in Q3/D11, so we need to scale them.
  int depth_shift = (TX_COEFF_DEPTH - 11) * 2;
  int depth_round = depth_shift > 1 ? (1 << depth_shift >> 1) : 0;
#else
156
  const int shift = av1_get_tx_scale(tx_size);
157
#endif
158 159
#if CONFIG_AOM_QM
  int seg_id = xd->mi[0]->mbmi.segment_id;
160
  // Use a flat matrix (i.e. no weighting) for 1D and Identity transforms
161 162 163
  const qm_val_t *iqmatrix = IS_2D_TRANSFORM(tx_type)
                                 ? pd->seg_iqmatrix[seg_id][tx_size]
                                 : cm->giqmatrix[NUM_QM_LEVELS - 1][0][tx_size];
164 165 166
#endif
#if CONFIG_NEW_QUANT
  int dq = get_dq_profile_from_ctx(mb->qindex, ctx, ref, plane_type);
167
  const dequant_val_type_nuq *dequant_val = p->dequant_val_nuq_QTX[dq];
168 169 170
#endif  // CONFIG_NEW_QUANT
  int64_t rd_cost0, rd_cost1;
  int16_t t0, t1;
171
  int i, final_eob = 0;
172
  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, xd->bd);
hui su's avatar
hui su committed
173 174 175 176
  int(*head_token_costs)[COEFF_CONTEXTS][TAIL_TOKENS] =
      mb->token_head_costs[txsize_sqr_map[tx_size]][plane_type][ref];
  int(*tail_token_costs)[COEFF_CONTEXTS][TAIL_TOKENS] =
      mb->token_tail_costs[txsize_sqr_map[tx_size]][plane_type][ref];
Dake He's avatar
Dake He committed
177

178
  const int64_t rdmult = (mb->rdmult * plane_rd_mult[ref][plane_type]) >> 1;
179
  int64_t rate0, rate1;
180 181 182 183 184
  int64_t eob_cost0, eob_cost1;
  tran_low_t before_best_eob_qc = 0;
  tran_low_t before_best_eob_dqc = 0;

  uint8_t token_cache[MAX_TX_SQUARE];
185 186 187 188

  // TODO(debargha): Implement a fast mode. For now just skip.
  if (fast_mode) return eob;

189 190
  for (i = 0; i < eob; i++) {
    const int rc = scan[i];
191
    token_cache[rc] = av1_pt_energy_class[av1_get_token(qcoeff[rc])];
192 193 194 195
  }

  /* Record the r-d cost */
  int64_t accu_rate = 0;
196 197 198
  // 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;
199
  rate0 = head_token_costs[0][ctx][0];
200
  int64_t best_block_rd_cost = RDCOST(rdmult, rate0, accu_error);
201 202

  // int64_t best_block_rd_cost_all0 = best_block_rd_cost;
hui su's avatar
hui su committed
203 204
  const int seg_eob =
      av1_get_tx_eob(&cm->seg, xd->mi[0]->mbmi.segment_id, tx_size);
205 206
  for (i = 0; i < eob; i++) {
    const int rc = scan[i];
207 208 209 210
    const int x = qcoeff[rc];
    const int sz = -(x < 0);
    const int band_cur = band_translate[i];
    const int ctx_cur = (i == 0) ? ctx : get_coef_context(nb, token_cache, i);
Yaowu Xu's avatar
Yaowu Xu committed
211
    const int8_t eob_val =
hui su's avatar
hui su committed
212 213
        (i + 1 == eob) ? (i + 1 == seg_eob ? LAST_EOB : EARLY_EOB) : NO_EOB;
    const int is_first = (i == 0);
214 215 216

    if (x == 0) {
      // no need to search when x == 0
217 218 219
      accu_rate += av1_get_coeff_token_cost(
          ZERO_TOKEN, eob_val, is_first, head_token_costs[band_cur][ctx_cur],
          tail_token_costs[band_cur][ctx_cur]);
220 221
      // accu_error does not change when x==0
    } else {
222 223 224 225 226 227 228 229
/*  Computing distortion
 */
// compute the distortion for the first candidate
// and the distortion for quantizing to 0.
#if CONFIG_DAALA_TX
      int dx0 = coeff[rc];
      const int64_t d0 = ((int64_t)dx0 * dx0 + depth_round) >> depth_shift;
#else
230
      int dx0 = abs(coeff[rc]) * (1 << shift);
Sebastien Alaiwan's avatar
Sebastien Alaiwan committed
231 232
      dx0 >>= xd->bd - 8;

233
      const int64_t d0 = (int64_t)dx0 * dx0;
234
#endif
235 236
      const int x_a = x - 2 * sz - 1;
      int dqv;
237
#if CONFIG_AOM_QM
238
      int iwt;
239
      dqv = dequant_ptr[rc != 0];
240 241 242 243
      if (iqmatrix != NULL) {
        iwt = iqmatrix[rc];
        dqv = ((iwt * (int)dqv) + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
      }
244 245 246 247
#else
      dqv = dequant_ptr[rc != 0];
#endif

248 249 250 251
#if CONFIG_DAALA_TX
      int dx = dqcoeff[rc] - coeff[rc];
      const int64_t d2 = ((int64_t)dx * dx + depth_round) >> depth_shift;
#else
252
      int dx = (dqcoeff[rc] - coeff[rc]) * (1 << shift);
253
      dx = signed_shift_right(dx, xd->bd - 8);
254
      const int64_t d2 = (int64_t)dx * dx;
255
#endif
256 257 258 259

      /* compute the distortion for the second candidate
       * x_a = x - 2 * sz + 1;
       */
260
      int64_t d2_a;
261
      if (x_a != 0) {
262 263 264 265 266 267 268 269 270
#if CONFIG_DAALA_TX
#if CONFIG_NEW_QUANT
        dx = av1_dequant_coeff_nuq(x, dqv, dequant_val[band_translate[i]]) -
             coeff[rc];
#else   // CONFIG_NEW_QUANT
        dx -= (dqv + sz) ^ sz;
#endif  // CONFIG_NEW_QUANT
        d2_a = ((int64_t)dx * dx + depth_round) >> depth_shift;
#else  // CONFIG_DAALA_TX
271 272
#if CONFIG_NEW_QUANT
        dx = av1_dequant_coeff_nuq(x, dqv, dequant_val[band_translate[i]]) -
Yaowu Xu's avatar
Yaowu Xu committed
273
             (coeff[rc] * (1 << shift));
Sebastien Alaiwan's avatar
Sebastien Alaiwan committed
274
        dx >>= xd->bd - 8;
275
#else   // CONFIG_NEW_QUANT
Sebastien Alaiwan's avatar
Sebastien Alaiwan committed
276
        dx -= ((dqv >> (xd->bd - 8)) + sz) ^ sz;
277 278
#endif  // CONFIG_NEW_QUANT
        d2_a = (int64_t)dx * dx;
279
#endif  // CONFIG_DAALA_TX
280 281 282
      } else {
        d2_a = d0;
      }
283

284
      // Computing RD cost
hui su's avatar
hui su committed
285
      int64_t base_bits;
286 287
      // rate cost of x
      base_bits = av1_get_token_cost(x, &t0, cat6_bits);
hui su's avatar
hui su committed
288 289 290 291
      rate0 = base_bits +
              av1_get_coeff_token_cost(t0, eob_val, is_first,
                                       head_token_costs[band_cur][ctx_cur],
                                       tail_token_costs[band_cur][ctx_cur]);
292
      // rate cost of x_a
293
      base_bits = av1_get_token_cost(x_a, &t1, cat6_bits);
hui su's avatar
hui su committed
294 295 296 297 298 299 300 301
      if (t1 == ZERO_TOKEN && eob_val) {
        rate1 = base_bits;
      } else {
        rate1 = base_bits +
                av1_get_coeff_token_cost(t1, eob_val, is_first,
                                         head_token_costs[band_cur][ctx_cur],
                                         tail_token_costs[band_cur][ctx_cur]);
      }
302

hui su's avatar
hui su committed
303
      int64_t next_bits0 = 0, next_bits1 = 0;
304
      if (i < eob - 1) {
hui su's avatar
hui su committed
305
        int ctx_next;
306 307
        const int band_next = band_translate[i + 1];
        const int token_next = av1_get_token(qcoeff[scan[i + 1]]);
Yaowu Xu's avatar
Yaowu Xu committed
308
        const int8_t eob_val_next =
hui su's avatar
hui su committed
309
            (i + 2 == eob) ? (i + 2 == seg_eob ? LAST_EOB : EARLY_EOB) : NO_EOB;
310 311 312

        token_cache[rc] = av1_pt_energy_class[t0];
        ctx_next = get_coef_context(nb, token_cache, i + 1);
313 314 315
        next_bits0 = av1_get_coeff_token_cost(
            token_next, eob_val_next, 0, head_token_costs[band_next][ctx_next],
            tail_token_costs[band_next][ctx_next]);
316 317 318

        token_cache[rc] = av1_pt_energy_class[t1];
        ctx_next = get_coef_context(nb, token_cache, i + 1);
319 320 321
        next_bits1 = av1_get_coeff_token_cost(
            token_next, eob_val_next, 0, head_token_costs[band_next][ctx_next],
            tail_token_costs[band_next][ctx_next]);
322 323
      }

324 325
      rd_cost0 = RDCOST(rdmult, (rate0 + next_bits0), d2);
      rd_cost1 = RDCOST(rdmult, (rate1 + next_bits1), d2_a);
326
      const int best_x = (rd_cost1 < rd_cost0);
327

328
      const int eob_v = (i + 1 == seg_eob) ? LAST_EOB : EARLY_EOB;
hui su's avatar
hui su committed
329 330 331 332 333 334 335
      int64_t next_eob_bits0, next_eob_bits1;
      int best_eob_x;
      next_eob_bits0 = av1_get_coeff_token_cost(
          t0, eob_v, is_first, head_token_costs[band_cur][ctx_cur],
          tail_token_costs[band_cur][ctx_cur]);
      eob_cost0 =
          RDCOST(rdmult, (accu_rate + next_eob_bits0), (accu_error + d2 - d0));
336 337
      eob_cost1 = eob_cost0;
      if (x_a != 0) {
hui su's avatar
hui su committed
338 339 340 341
        next_eob_bits1 = av1_get_coeff_token_cost(
            t1, eob_v, is_first, head_token_costs[band_cur][ctx_cur],
            tail_token_costs[band_cur][ctx_cur]);
        eob_cost1 = RDCOST(rdmult, (accu_rate + next_eob_bits1),
342 343 344 345 346 347
                           (accu_error + d2_a - d0));
        best_eob_x = (eob_cost1 < eob_cost0);
      } else {
        best_eob_x = 0;
      }

348 349 350
      const int dqc = dqcoeff[rc];
      int dqc_a = 0;
      if (best_x || best_eob_x) {
351
        if (x_a != 0) {
352 353 354 355 356 357 358 359 360
#if CONFIG_DAALA_TX
#if CONFIG_NEW_QUANT
          dqc_a = av1_dequant_abscoeff_nuq(abs(x_a), dqv,
                                           dequant_val[band_translate[i]]);
          if (sz) dqc_a = -dqc_a;
#else
          dqc_a = x_a * dqv;
#endif  // CONFIG_NEW_QUANT
#else   // CONFIG_DAALA_TX
361 362 363 364 365 366
#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
367 368
          if (x_a < 0)
            dqc_a = -((-x_a * dqv) >> shift);
369
          else
370
            dqc_a = (x_a * dqv) >> shift;
371
#endif  // CONFIG_NEW_QUANT
372
#endif  // CONFIG_DAALA_TX
373 374
        } else {
          dqc_a = 0;
375
        }
376 377 378 379
      }

      // record the better quantized value
      if (best_x) {
380
        assert(d2_a <= d0);
381 382 383 384 385 386
        qcoeff[rc] = x_a;
        dqcoeff[rc] = dqc_a;
        accu_rate += rate1;
        accu_error += d2_a - d0;
        token_cache[rc] = av1_pt_energy_class[t1];
      } else {
387
        assert(d2 <= d0);
388 389 390 391
        accu_rate += rate0;
        accu_error += d2 - d0;
        token_cache[rc] = av1_pt_energy_class[t0];
      }
392
      assert(accu_error >= 0);
393 394

      // determine whether to move the eob position to i+1
395 396
      const int use_a = (x_a != 0) && (best_eob_x);
      const int64_t best_eob_cost_i = use_a ? eob_cost1 : eob_cost0;
397 398 399
      if (best_eob_cost_i < best_block_rd_cost) {
        best_block_rd_cost = best_eob_cost_i;
        final_eob = i + 1;
400 401 402 403 404 405 406
        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;
        }
407 408 409 410 411 412
      }
    }  // if (x==0)
  }    // for (i)

  assert(final_eob <= eob);
  if (final_eob > 0) {
413
    assert(before_best_eob_qc != 0);
414 415
    i = final_eob - 1;
    int rc = scan[i];
416 417
    qcoeff[rc] = before_best_eob_qc;
    dqcoeff[rc] = before_best_eob_dqc;
418 419 420 421 422 423 424 425
  }

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

426
  p->eobs[block] = final_eob;
427 428
  return final_eob;
}
429
#endif  // !CONFIG_LV_MAP
430

Jingning Han's avatar
Jingning Han committed
431 432 433
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,
434
                   const ENTROPY_CONTEXT *l, int fast_mode) {
435 436 437 438 439 440
  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
441

442
#if !CONFIG_LV_MAP
443
  (void)plane_bsize;
Jingning Han's avatar
Jingning Han committed
444 445
  (void)blk_row;
  (void)blk_col;
446
  int ctx = get_entropy_context(tx_size, a, l);
447 448
  return optimize_b_greedy(cm, mb, plane, blk_row, blk_col, block, tx_size, ctx,
                           fast_mode);
449 450 451
#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
452
  return av1_optimize_txb(cm, mb, plane, blk_row, blk_col, block, tx_size,
453
                          &txb_ctx, fast_mode);
454
#endif  // !CONFIG_LV_MAP
455
}
456

Angie Chiang's avatar
Angie Chiang committed
457 458
typedef enum QUANT_FUNC {
  QUANT_FUNC_LOWBD = 0,
459
  QUANT_FUNC_HIGHBD = 1,
460
  QUANT_FUNC_TYPES = 2
Angie Chiang's avatar
Angie Chiang committed
461 462
} QUANT_FUNC;

463 464
static AV1_QUANT_FACADE
    quant_func_list[AV1_XFORM_QUANT_TYPES][QUANT_FUNC_TYPES] = {
465
#if !CONFIG_NEW_QUANT
466
      { av1_quantize_fp_facade, av1_highbd_quantize_fp_facade },
Yaowu Xu's avatar
Yaowu Xu committed
467 468
      { av1_quantize_b_facade, av1_highbd_quantize_b_facade },
      { av1_quantize_dc_facade, av1_highbd_quantize_dc_facade },
469
#else   // !CONFIG_NEW_QUANT
470 471 472
      { 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 },
473
#endif  // !CONFIG_NEW_QUANT
474 475
      { NULL, NULL }
    };
476

Yushin Cho's avatar
Yushin Cho committed
477
#if !CONFIG_TXMG
478
typedef void (*fwdTxfmFunc)(const int16_t *diff, tran_low_t *coeff, int stride,
479
                            TxfmParam *txfm_param);
480 481
static const fwdTxfmFunc fwd_txfm_func[2] = { av1_fwd_txfm,
                                              av1_highbd_fwd_txfm };
Angie Chiang's avatar
Angie Chiang committed
482
#endif
483

Angie Chiang's avatar
Angie Chiang committed
484 485
void av1_xform_quant(const AV1_COMMON *cm, MACROBLOCK *x, int plane, int block,
                     int blk_row, int blk_col, BLOCK_SIZE plane_bsize,
486 487
                     TX_SIZE tx_size, int ctx,
                     AV1_XFORM_QUANT xform_quant_idx) {
Jingning Han's avatar
Jingning Han committed
488
  MACROBLOCKD *const xd = &x->e_mbd;
489
  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
Yushin Cho's avatar
Yushin Cho committed
490
#if !CONFIG_DIST_8X8
Jingning Han's avatar
Jingning Han committed
491 492
  const struct macroblock_plane *const p = &x->plane[plane];
  const struct macroblockd_plane *const pd = &xd->plane[plane];
493 494 495 496
#else
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
#endif
497
  PLANE_TYPE plane_type = get_plane_type(plane);
498 499
  TX_TYPE tx_type =
      av1_get_tx_type(plane_type, xd, blk_row, blk_col, block, tx_size);
Angie Chiang's avatar
Angie Chiang committed
500

501
#if CONFIG_NEW_QUANT
502
  const int is_inter = is_inter_block(mbmi);
Angie Chiang's avatar
Angie Chiang committed
503 504
#endif
  const SCAN_ORDER *const scan_order = get_scan(cm, tx_size, tx_type, mbmi);
Jingning Han's avatar
Jingning Han committed
505 506 507 508
  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];
509
  const int diff_stride = block_size_wide[plane_bsize];
Yushin Cho's avatar
Yushin Cho committed
510
#if CONFIG_AOM_QM
511
  int seg_id = mbmi->segment_id;
512
  // Use a flat matrix (i.e. no weighting) for 1D and Identity transforms
513 514 515 516 517 518
  const qm_val_t *qmatrix = IS_2D_TRANSFORM(tx_type)
                                ? pd->seg_qmatrix[seg_id][tx_size]
                                : cm->gqmatrix[NUM_QM_LEVELS - 1][0][tx_size];
  const qm_val_t *iqmatrix = IS_2D_TRANSFORM(tx_type)
                                 ? pd->seg_iqmatrix[seg_id][tx_size]
                                 : cm->giqmatrix[NUM_QM_LEVELS - 1][0][tx_size];
519
#endif
Angie Chiang's avatar
Angie Chiang committed
520

521
  TxfmParam txfm_param;
522

523
#if CONFIG_DIST_8X8 || CONFIG_MRC_TX
Yushin Cho's avatar
Yushin Cho committed
524 525
  uint8_t *dst;
  const int dst_stride = pd->dst.stride;
Yushin Cho's avatar
Yushin Cho committed
526
#if CONFIG_DIST_8X8
Lester Lu's avatar
Lester Lu committed
527
  int16_t *pred;
528 529
  const int txw = tx_size_wide[tx_size];
  const int txh = tx_size_high[tx_size];
Yushin Cho's avatar
Yushin Cho committed
530 531
  int i, j;
#endif
Lester Lu's avatar
Lester Lu committed
532
#endif
Yushin Cho's avatar
Yushin Cho committed
533

534
  const int tx2d_size = tx_size_2d[tx_size];
535
  QUANT_PARAM qparam;
536 537
  const int16_t *src_diff;

538 539
  src_diff =
      &p->src_diff[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
540
  qparam.log_scale = av1_get_tx_scale(tx_size);
541 542 543 544 545 546 547 548
#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
Yushin Cho's avatar
Yushin Cho committed
549

550
#if CONFIG_DIST_8X8 || CONFIG_MRC_TX
Yushin Cho's avatar
Yushin Cho committed
551
  dst = &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
552
#endif  // CONFIG_DIST_8X8 ||
553
        // CONFIG_MRC_TX
554 555

#if CONFIG_DIST_8X8
Yushin Cho's avatar
Yushin Cho committed
556
  if (x->using_dist_8x8) {
557
    pred = &pd->pred[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
558

Thomas Daede's avatar
Thomas Daede committed
559 560
// copy uint8 orig and predicted block to int16 buffer
// in order to use existing VP10 transform functions
561
#if CONFIG_HIGHBITDEPTH
562 563 564 565 566 567
    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
      for (j = 0; j < txh; j++)
        for (i = 0; i < txw; i++)
          pred[diff_stride * j + i] =
              CONVERT_TO_SHORTPTR(dst)[dst_stride * j + i];
    } else {
568
#endif  // CONFIG_HIGHBITDEPTH
569 570 571
      for (j = 0; j < txh; j++)
        for (i = 0; i < txw; i++)
          pred[diff_stride * j + i] = dst[dst_stride * j + i];
572
#if CONFIG_HIGHBITDEPTH
573
    }
574
#endif  // CONFIG_HIGHBITDEPTH
575
  }
Yushin Cho's avatar
Yushin Cho committed
576
#endif  // CONFIG_DIST_8X8
Yushin Cho's avatar
Yushin Cho committed
577

578
  (void)ctx;
579

580 581 582
  txfm_param.tx_type = tx_type;
  txfm_param.tx_size = tx_size;
  txfm_param.lossless = xd->lossless[mbmi->segment_id];
583 584 585
  txfm_param.tx_set_type =
      get_ext_tx_set_type(txfm_param.tx_size, plane_bsize, is_inter_block(mbmi),
                          cm->reduced_tx_set_used);
Sebastien Alaiwan's avatar
Sebastien Alaiwan committed
586
#if CONFIG_MRC_TX
587
  txfm_param.is_inter = is_inter_block(mbmi);
588
#endif
589
#if CONFIG_MRC_TX
590 591
  txfm_param.dst = dst;
  txfm_param.stride = dst_stride;
592 593
#if CONFIG_MRC_TX
  txfm_param.valid_mask = &mbmi->valid_mrc_mask;
Sarah Parker's avatar
Sarah Parker committed
594 595 596
#if SIGNAL_ANY_MRC_MASK
  txfm_param.mask = BLOCK_OFFSET(xd->mrc_mask, block);
#endif  // SIGNAL_ANY_MRC_MASK
597
#endif  // CONFIG_MRC_TX
598
#endif  // CONFIG_MRC_TX
Angie Chiang's avatar
Angie Chiang committed
599

600
  txfm_param.bd = xd->bd;
601
  txfm_param.is_hbd = get_bitdepth_data_path_index(xd);
Angie Chiang's avatar
Angie Chiang committed
602 603 604 605

#if CONFIG_TXMG
  av1_highbd_fwd_txfm(src_diff, coeff, diff_stride, &txfm_param);
#else   // CONFIG_TXMG
606
  fwd_txfm_func[txfm_param.is_hbd](src_diff, coeff, diff_stride, &txfm_param);
Angie Chiang's avatar
Angie Chiang committed
607
#endif  // CONFIG_TXMG
608

Yaowu Xu's avatar
Yaowu Xu committed
609
  if (xform_quant_idx != AV1_XFORM_QUANT_SKIP_QUANT) {
610
    if (LIKELY(!x->skip_block)) {
611 612 613 614
#if CONFIG_DAALA_TX
      quant_func_list[xform_quant_idx][1](coeff, tx2d_size, p, qcoeff, dqcoeff,
                                          eob, scan_order, &qparam);
#else
615
      quant_func_list[xform_quant_idx][txfm_param.is_hbd](
616
          coeff, tx2d_size, p, qcoeff, dqcoeff, eob, scan_order, &qparam);
617
#endif
618
    } else {
Yaowu Xu's avatar
Yaowu Xu committed
619
      av1_quantize_skip(tx2d_size, qcoeff, dqcoeff, eob);
Jingning Han's avatar
Jingning Han committed
620 621
    }
  }
622 623 624 625
#if CONFIG_LV_MAP
  p->txb_entropy_ctx[block] =
      (uint8_t)av1_get_txb_entropy_context(qcoeff, scan_order, *eob);
#endif  // CONFIG_LV_MAP
626
  return;
Jingning Han's avatar
Jingning Han committed
627 628
}

629
static void encode_block(int plane, int block, int blk_row, int blk_col,
630
                         BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
Jingning Han's avatar
Jingning Han committed
631
  struct encode_b_args *const args = arg;
Angie Chiang's avatar
Angie Chiang committed
632
  AV1_COMMON *cm = args->cm;
Jingning Han's avatar
Jingning Han committed
633 634
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
635
  int ctx;
Jingning Han's avatar
Jingning Han committed
636 637 638
  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);
Sarah Parker's avatar
Sarah Parker committed
639 640 641
#if CONFIG_MRC_TX && SIGNAL_ANY_MRC_MASK
  uint8_t *mrc_mask = BLOCK_OFFSET(xd->mrc_mask, block);
#endif  // CONFIG_MRC_TX && SIGNAL_ANY_MRC_MASK
Jingning Han's avatar
Jingning Han committed
642 643
  uint8_t *dst;
  ENTROPY_CONTEXT *a, *l;
Yushin Cho's avatar
Yushin Cho committed
644

645
  int bw = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
646 647
  dst = &pd->dst
             .buf[(blk_row * pd->dst.stride + blk_col) << tx_size_wide_log2[0]];
648

649 650
  a = &args->ta[blk_col];
  l = &args->tl[blk_row];
651
  ctx = get_entropy_context(tx_size, a, l);
Jingning Han's avatar
Jingning Han committed
652

653
  // Assert not magic number (uninitialized).
654
  assert(x->blk_skip[plane][blk_row * bw + blk_col] != 234);
655

656
  if (x->blk_skip[plane][blk_row * bw + blk_col] == 0) {
Angie Chiang's avatar
Angie Chiang committed
657
    av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
658
                    ctx, AV1_XFORM_QUANT_FP);
659
  } else {
660
    p->eobs[block] = 0;
661
  }
662

Jingning Han's avatar
Jingning Han committed
663
  av1_optimize_b(cm, x, plane, blk_row, blk_col, block, plane_bsize, tx_size, a,
664
                 l, CONFIG_LV_MAP);
665

666 667
  av1_set_txb_context(x, plane, block, tx_size, a, l);

668
  if (p->eobs[block]) *(args->skip) = 0;
Jingning Han's avatar
Jingning Han committed
669

670
  if (p->eobs[block] != 0)
671

672 673 674 675
  {
    TX_TYPE tx_type =
        av1_get_tx_type(pd->plane_type, xd, blk_row, blk_col, block, tx_size);
    av1_inverse_transform_block(xd, dqcoeff,
Sarah Parker's avatar
Sarah Parker committed
676
#if CONFIG_MRC_TX && SIGNAL_ANY_MRC_MASK
677
                                mrc_mask,
Sarah Parker's avatar
Sarah Parker committed
678
#endif  // CONFIG_MRC_TX && SIGNAL_ANY_MRC_MASK
679
                                plane, tx_type, tx_size, dst, pd->dst.stride,
680 681
                                p->eobs[block]);
  }
Jingning Han's avatar
Jingning Han committed
682 683
}

684 685 686 687 688 689 690
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;
691
  const BLOCK_SIZE bsize = txsize_to_bsize[tx_size];
692
  const struct macroblockd_plane *const pd = &xd->plane[plane];
693 694
  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
695
  TX_SIZE plane_tx_size;
696 697
  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
698

699
  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
700

701 702 703
  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
704

705
  if (tx_size == plane_tx_size) {
706
    encode_block(plane, block, blk_row, blk_col, plane_bsize, tx_size, arg);
707
  } else {
708
    assert(tx_size < TX_SIZES_ALL);
709
    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
710 711
    assert(IMPLIES(tx_size <= TX_4X4, sub_txs == tx_size));
    assert(IMPLIES(tx_size > TX_4X4, sub_txs < tx_size));
712
    // This is the square transform block partition entry point.
713 714 715 716
    const int bsw = tx_size_wide_unit[sub_txs];
    const int bsh = tx_size_high_unit[sub_txs];
    const int step = bsh * bsw;
    assert(bsw > 0 && bsh > 0);
717

718 719 720 721
    for (int row = 0; row < tx_size_high_unit[tx_size]; row += bsh) {
      for (int col = 0; col < tx_size_wide_unit[tx_size]; col += bsw) {
        const int offsetr = blk_row + row;
        const int offsetc = blk_col + col;
722

723
        if (offsetr >= max_blocks_high || offsetc >= max_blocks_wide) continue;
724

725 726 727 728
        encode_block_inter(plane, block, offsetr, offsetc, plane_bsize, sub_txs,
                           arg);
        block += step;
      }
729 730 731 732
    }
  }
}

Angie Chiang's avatar
Angie Chiang committed
733 734 735 736 737
typedef struct encode_block_pass1_args {
  AV1_COMMON *cm;
  MACROBLOCK *x;
} encode_block_pass1_args;

738
static void encode_block_pass1(int plane, int block, int blk_row, int blk_col,
739 740
                               BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
                               void *arg) {
Angie Chiang's avatar
Angie Chiang committed
741 742 743
  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
744 745 746 747
  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);
748
  TxfmParam txfm_param;
Jingning Han's avatar
Jingning Han committed
749
  uint8_t *dst;
750
  int ctx = 0;
751 752
  dst = &pd->dst
             .buf[(blk_row * pd->dst.stride + blk_col) << tx_size_wide_log2[0]];
Jingning Han's avatar
Jingning Han committed
753

Angie Chiang's avatar
Angie Chiang committed
754
  av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
755
                  ctx, AV1_XFORM_QUANT_B);
756

Yushin Cho's avatar
Yushin Cho committed
757
  if (p->eobs[block] > 0) {
758
    txfm_param.bd = xd->bd;
759
    txfm_param.is_hbd = get_bitdepth_data_path_index(xd);
760
    txfm_param.tx_type = DCT_DCT;
761
    txfm_param.tx_size = tx_size;
762 763
    txfm_param.eob = p->eobs[block];
    txfm_param.lossless = xd->lossless[xd->mi[0]->mbmi.segment_id];
764 765 766
    txfm_param.tx_set_type = get_ext_tx_set_type(
        txfm_param.tx_size, plane_bsize, is_inter_block(&xd->mi[0]->mbmi),
        cm->reduced_tx_set_used);
767 768 769
#if CONFIG_DAALA_TX
    daala_inv_txfm_add(dqcoeff, dst, pd->dst.stride, &txfm_param);
#else
770
#if CONFIG_HIGHBITDEPTH
771
    if (txfm_param.is_hbd) {
772
      av1_highbd_inv_txfm_add_4x4(dqcoeff, dst, pd->dst.stride, &txfm_param);
773
      return;
Jingning Han's avatar
Jingning Han committed
774
    }
775
#endif  //  CONFIG_HIGHBITDEPTH