encodemb.c 33.3 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
#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
  const int shift = av1_get_tx_scale(tx_size);
151 152
#if CONFIG_AOM_QM
  int seg_id = xd->mi[0]->mbmi.segment_id;
153
  // Use a flat matrix (i.e. no weighting) for 1D and Identity transforms
154 155 156
  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];
157 158 159
#endif
#if CONFIG_NEW_QUANT
  int dq = get_dq_profile_from_ctx(mb->qindex, ctx, ref, plane_type);
160
  const dequant_val_type_nuq *dequant_val = p->dequant_val_nuq_QTX[dq];
161 162 163
#endif  // CONFIG_NEW_QUANT
  int64_t rd_cost0, rd_cost1;
  int16_t t0, t1;
164
  int i, final_eob = 0;
165
  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, xd->bd);
hui su's avatar
hui su committed
166 167 168 169
  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
170

171
  const int64_t rdmult = (mb->rdmult * plane_rd_mult[ref][plane_type]) >> 1;
172
  int64_t rate0, rate1;
173 174 175 176 177
  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];
178 179 180 181

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

182 183
  for (i = 0; i < eob; i++) {
    const int rc = scan[i];
184
    token_cache[rc] = av1_pt_energy_class[av1_get_token(qcoeff[rc])];
185 186 187 188
  }

  /* Record the r-d cost */
  int64_t accu_rate = 0;
189 190 191
  // 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;
192
  rate0 = head_token_costs[0][ctx][0];
193
  int64_t best_block_rd_cost = RDCOST(rdmult, rate0, accu_error);
194 195

  // int64_t best_block_rd_cost_all0 = best_block_rd_cost;
hui su's avatar
hui su committed
196 197
  const int seg_eob =
      av1_get_tx_eob(&cm->seg, xd->mi[0]->mbmi.segment_id, tx_size);
198 199
  for (i = 0; i < eob; i++) {
    const int rc = scan[i];
200 201 202 203
    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
204
    const int8_t eob_val =
hui su's avatar
hui su committed
205 206
        (i + 1 == eob) ? (i + 1 == seg_eob ? LAST_EOB : EARLY_EOB) : NO_EOB;
    const int is_first = (i == 0);
207 208 209

    if (x == 0) {
      // no need to search when x == 0
210 211 212
      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]);
213 214 215 216 217 218
      // 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.
219
      int dx0 = abs(coeff[rc]) * (1 << shift);
Sebastien Alaiwan's avatar
Sebastien Alaiwan committed
220 221
      dx0 >>= xd->bd - 8;

222 223 224
      const int64_t d0 = (int64_t)dx0 * dx0;
      const int x_a = x - 2 * sz - 1;
      int dqv;
225
#if CONFIG_AOM_QM
226
      int iwt;
227
      dqv = dequant_ptr[rc != 0];
228 229 230 231
      if (iqmatrix != NULL) {
        iwt = iqmatrix[rc];
        dqv = ((iwt * (int)dqv) + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
      }
232 233 234 235
#else
      dqv = dequant_ptr[rc != 0];
#endif

236
      int dx = (dqcoeff[rc] - coeff[rc]) * (1 << shift);
237
      dx = signed_shift_right(dx, xd->bd - 8);
238
      const int64_t d2 = (int64_t)dx * dx;
239 240 241 242

      /* compute the distortion for the second candidate
       * x_a = x - 2 * sz + 1;
       */
243
      int64_t d2_a;
244 245 246
      if (x_a != 0) {
#if CONFIG_NEW_QUANT
        dx = av1_dequant_coeff_nuq(x, dqv, dequant_val[band_translate[i]]) -
Yaowu Xu's avatar
Yaowu Xu committed
247
             (coeff[rc] * (1 << shift));
Sebastien Alaiwan's avatar
Sebastien Alaiwan committed
248
        dx >>= xd->bd - 8;
249
#else   // CONFIG_NEW_QUANT
Sebastien Alaiwan's avatar
Sebastien Alaiwan committed
250
        dx -= ((dqv >> (xd->bd - 8)) + sz) ^ sz;
251 252 253 254 255
#endif  // CONFIG_NEW_QUANT
        d2_a = (int64_t)dx * dx;
      } else {
        d2_a = d0;
      }
256
      // Computing RD cost
hui su's avatar
hui su committed
257
      int64_t base_bits;
258 259
      // rate cost of x
      base_bits = av1_get_token_cost(x, &t0, cat6_bits);
hui su's avatar
hui su committed
260 261 262 263
      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]);
264
      // rate cost of x_a
265
      base_bits = av1_get_token_cost(x_a, &t1, cat6_bits);
hui su's avatar
hui su committed
266 267 268 269 270 271 272 273
      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]);
      }
274

hui su's avatar
hui su committed
275
      int64_t next_bits0 = 0, next_bits1 = 0;
276
      if (i < eob - 1) {
hui su's avatar
hui su committed
277
        int ctx_next;
278 279
        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
280
        const int8_t eob_val_next =
hui su's avatar
hui su committed
281
            (i + 2 == eob) ? (i + 2 == seg_eob ? LAST_EOB : EARLY_EOB) : NO_EOB;
282 283 284

        token_cache[rc] = av1_pt_energy_class[t0];
        ctx_next = get_coef_context(nb, token_cache, i + 1);
285 286 287
        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]);
288 289 290

        token_cache[rc] = av1_pt_energy_class[t1];
        ctx_next = get_coef_context(nb, token_cache, i + 1);
291 292 293
        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]);
294 295
      }

296 297
      rd_cost0 = RDCOST(rdmult, (rate0 + next_bits0), d2);
      rd_cost1 = RDCOST(rdmult, (rate1 + next_bits1), d2_a);
298
      const int best_x = (rd_cost1 < rd_cost0);
299

300
      const int eob_v = (i + 1 == seg_eob) ? LAST_EOB : EARLY_EOB;
hui su's avatar
hui su committed
301 302 303 304 305 306 307
      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));
308 309
      eob_cost1 = eob_cost0;
      if (x_a != 0) {
hui su's avatar
hui su committed
310 311 312 313
        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),
314 315 316 317 318 319
                           (accu_error + d2_a - d0));
        best_eob_x = (eob_cost1 < eob_cost0);
      } else {
        best_eob_x = 0;
      }

320 321 322
      const int dqc = dqcoeff[rc];
      int dqc_a = 0;
      if (best_x || best_eob_x) {
323 324 325 326 327 328 329
        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
330 331
          if (x_a < 0)
            dqc_a = -((-x_a * dqv) >> shift);
332
          else
333
            dqc_a = (x_a * dqv) >> shift;
334 335 336 337 338 339 340 341
#endif  // CONFIG_NEW_QUANT
        } else {
          dqc_a = 0;
        }  // if (x_a != 0)
      }

      // record the better quantized value
      if (best_x) {
342
        assert(d2_a <= d0);
343 344 345 346 347 348
        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 {
349
        assert(d2 <= d0);
350 351 352 353
        accu_rate += rate0;
        accu_error += d2 - d0;
        token_cache[rc] = av1_pt_energy_class[t0];
      }
354
      assert(accu_error >= 0);
355 356

      // determine whether to move the eob position to i+1
357 358
      const int use_a = (x_a != 0) && (best_eob_x);
      const int64_t best_eob_cost_i = use_a ? eob_cost1 : eob_cost0;
359 360 361
      if (best_eob_cost_i < best_block_rd_cost) {
        best_block_rd_cost = best_eob_cost_i;
        final_eob = i + 1;
362 363 364 365 366 367 368
        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;
        }
369 370 371 372 373 374
      }
    }  // if (x==0)
  }    // for (i)

  assert(final_eob <= eob);
  if (final_eob > 0) {
375
    assert(before_best_eob_qc != 0);
376 377
    i = final_eob - 1;
    int rc = scan[i];
378 379
    qcoeff[rc] = before_best_eob_qc;
    dqcoeff[rc] = before_best_eob_dqc;
380 381 382 383 384 385 386 387
  }

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

388
  p->eobs[block] = final_eob;
389 390
  return final_eob;
}
391
#endif  // !CONFIG_LV_MAP
392

Jingning Han's avatar
Jingning Han committed
393 394 395
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,
396
                   const ENTROPY_CONTEXT *l, int fast_mode) {
397 398 399 400 401 402
  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
403

404
#if !CONFIG_LV_MAP
405
  (void)plane_bsize;
Jingning Han's avatar
Jingning Han committed
406 407
  (void)blk_row;
  (void)blk_col;
408
  int ctx = get_entropy_context(tx_size, a, l);
409 410
  return optimize_b_greedy(cm, mb, plane, blk_row, blk_col, block, tx_size, ctx,
                           fast_mode);
411 412 413
#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
414
  return av1_optimize_txb(cm, mb, plane, blk_row, blk_col, block, tx_size,
415
                          &txb_ctx, fast_mode);
416
#endif  // !CONFIG_LV_MAP
417
}
418

Angie Chiang's avatar
Angie Chiang committed
419 420
typedef enum QUANT_FUNC {
  QUANT_FUNC_LOWBD = 0,
421
  QUANT_FUNC_HIGHBD = 1,
422
  QUANT_FUNC_TYPES = 2
Angie Chiang's avatar
Angie Chiang committed
423 424
} QUANT_FUNC;

425 426
static AV1_QUANT_FACADE
    quant_func_list[AV1_XFORM_QUANT_TYPES][QUANT_FUNC_TYPES] = {
427
#if !CONFIG_NEW_QUANT
428
      { av1_quantize_fp_facade, av1_highbd_quantize_fp_facade },
Yaowu Xu's avatar
Yaowu Xu committed
429 430
      { av1_quantize_b_facade, av1_highbd_quantize_b_facade },
      { av1_quantize_dc_facade, av1_highbd_quantize_dc_facade },
431
#else   // !CONFIG_NEW_QUANT
432 433 434
      { 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 },
435
#endif  // !CONFIG_NEW_QUANT
436 437
      { NULL, NULL }
    };
438

Yushin Cho's avatar
Yushin Cho committed
439
#if !CONFIG_TXMG
440
typedef void (*fwdTxfmFunc)(const int16_t *diff, tran_low_t *coeff, int stride,
441
                            TxfmParam *txfm_param);
442 443
static const fwdTxfmFunc fwd_txfm_func[2] = { av1_fwd_txfm,
                                              av1_highbd_fwd_txfm };
Angie Chiang's avatar
Angie Chiang committed
444
#endif
445

Angie Chiang's avatar
Angie Chiang committed
446 447
void av1_xform_quant(const AV1_COMMON *cm, MACROBLOCK *x, int plane, int block,
                     int blk_row, int blk_col, BLOCK_SIZE plane_bsize,
448 449
                     TX_SIZE tx_size, int ctx,
                     AV1_XFORM_QUANT xform_quant_idx) {
Jingning Han's avatar
Jingning Han committed
450
  MACROBLOCKD *const xd = &x->e_mbd;
451
  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
Yushin Cho's avatar
Yushin Cho committed
452
#if !CONFIG_DIST_8X8
Jingning Han's avatar
Jingning Han committed
453 454
  const struct macroblock_plane *const p = &x->plane[plane];
  const struct macroblockd_plane *const pd = &xd->plane[plane];
455 456 457 458
#else
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
#endif
459
  PLANE_TYPE plane_type = get_plane_type(plane);
460 461
  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
462

463
#if CONFIG_NEW_QUANT
464
  const int is_inter = is_inter_block(mbmi);
Angie Chiang's avatar
Angie Chiang committed
465 466
#endif
  const SCAN_ORDER *const scan_order = get_scan(cm, tx_size, tx_type, mbmi);
Jingning Han's avatar
Jingning Han committed
467 468 469 470
  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];
471
  const int diff_stride = block_size_wide[plane_bsize];
Yushin Cho's avatar
Yushin Cho committed
472
#if CONFIG_AOM_QM
473
  int seg_id = mbmi->segment_id;
474
  // Use a flat matrix (i.e. no weighting) for 1D and Identity transforms
475 476 477 478 479 480
  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];
481
#endif
Angie Chiang's avatar
Angie Chiang committed
482

483
  TxfmParam txfm_param;
484

485
#if CONFIG_DIST_8X8 || CONFIG_MRC_TX
Yushin Cho's avatar
Yushin Cho committed
486 487
  uint8_t *dst;
  const int dst_stride = pd->dst.stride;
Yushin Cho's avatar
Yushin Cho committed
488
#if CONFIG_DIST_8X8
Lester Lu's avatar
Lester Lu committed
489
  int16_t *pred;
490 491
  const int txw = tx_size_wide[tx_size];
  const int txh = tx_size_high[tx_size];
Yushin Cho's avatar
Yushin Cho committed
492 493
  int i, j;
#endif
Lester Lu's avatar
Lester Lu committed
494
#endif
Yushin Cho's avatar
Yushin Cho committed
495

496
  const int tx2d_size = tx_size_2d[tx_size];
497
  QUANT_PARAM qparam;
498 499
  const int16_t *src_diff;

500 501
  src_diff =
      &p->src_diff[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
502
  qparam.log_scale = av1_get_tx_scale(tx_size);
503 504 505 506 507 508 509 510
#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
511

512
#if CONFIG_DIST_8X8 || CONFIG_MRC_TX
Yushin Cho's avatar
Yushin Cho committed
513
  dst = &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
514
#endif  // CONFIG_DIST_8X8 ||
515
        // CONFIG_MRC_TX
516 517

#if CONFIG_DIST_8X8
Yushin Cho's avatar
Yushin Cho committed
518
  if (x->using_dist_8x8) {
519
    pred = &pd->pred[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
520

Thomas Daede's avatar
Thomas Daede committed
521 522
// copy uint8 orig and predicted block to int16 buffer
// in order to use existing VP10 transform functions
523
#if CONFIG_HIGHBITDEPTH
524 525 526 527 528 529
    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 {
530
#endif  // CONFIG_HIGHBITDEPTH
531 532 533
      for (j = 0; j < txh; j++)
        for (i = 0; i < txw; i++)
          pred[diff_stride * j + i] = dst[dst_stride * j + i];
534
#if CONFIG_HIGHBITDEPTH
535
    }
536
#endif  // CONFIG_HIGHBITDEPTH
537
  }
Yushin Cho's avatar
Yushin Cho committed
538
#endif  // CONFIG_DIST_8X8
Yushin Cho's avatar
Yushin Cho committed
539

540
  (void)ctx;
541

542 543 544
  txfm_param.tx_type = tx_type;
  txfm_param.tx_size = tx_size;
  txfm_param.lossless = xd->lossless[mbmi->segment_id];
545 546 547
  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
548
#if CONFIG_MRC_TX
549
  txfm_param.is_inter = is_inter_block(mbmi);
550
#endif
551
#if CONFIG_MRC_TX
552 553
  txfm_param.dst = dst;
  txfm_param.stride = dst_stride;
554 555
#if CONFIG_MRC_TX
  txfm_param.valid_mask = &mbmi->valid_mrc_mask;
Sarah Parker's avatar
Sarah Parker committed
556 557 558
#if SIGNAL_ANY_MRC_MASK
  txfm_param.mask = BLOCK_OFFSET(xd->mrc_mask, block);
#endif  // SIGNAL_ANY_MRC_MASK
559
#endif  // CONFIG_MRC_TX
560
#endif  // CONFIG_MRC_TX
Angie Chiang's avatar
Angie Chiang committed
561

562
  txfm_param.bd = xd->bd;
563
  txfm_param.is_hbd = get_bitdepth_data_path_index(xd);
Angie Chiang's avatar
Angie Chiang committed
564 565 566 567

#if CONFIG_TXMG
  av1_highbd_fwd_txfm(src_diff, coeff, diff_stride, &txfm_param);
#else   // CONFIG_TXMG
568
  fwd_txfm_func[txfm_param.is_hbd](src_diff, coeff, diff_stride, &txfm_param);
Angie Chiang's avatar
Angie Chiang committed
569
#endif  // CONFIG_TXMG
570

Yaowu Xu's avatar
Yaowu Xu committed
571
  if (xform_quant_idx != AV1_XFORM_QUANT_SKIP_QUANT) {
572
    if (LIKELY(!x->skip_block)) {
573
      quant_func_list[xform_quant_idx][txfm_param.is_hbd](
574
          coeff, tx2d_size, p, qcoeff, dqcoeff, eob, scan_order, &qparam);
575
    } else {
Yaowu Xu's avatar
Yaowu Xu committed
576
      av1_quantize_skip(tx2d_size, qcoeff, dqcoeff, eob);
Jingning Han's avatar
Jingning Han committed
577 578
    }
  }
579 580 581 582
#if CONFIG_LV_MAP
  p->txb_entropy_ctx[block] =
      (uint8_t)av1_get_txb_entropy_context(qcoeff, scan_order, *eob);
#endif  // CONFIG_LV_MAP
583
  return;
Jingning Han's avatar
Jingning Han committed
584 585
}

586
static void encode_block(int plane, int block, int blk_row, int blk_col,
587
                         BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
Jingning Han's avatar
Jingning Han committed
588
  struct encode_b_args *const args = arg;
Angie Chiang's avatar
Angie Chiang committed
589
  AV1_COMMON *cm = args->cm;
Jingning Han's avatar
Jingning Han committed
590 591
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
592
  int ctx;
Jingning Han's avatar
Jingning Han committed
593 594 595
  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
596 597 598
#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
599 600
  uint8_t *dst;
  ENTROPY_CONTEXT *a, *l;
Yushin Cho's avatar
Yushin Cho committed
601

602
  int bw = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
603 604
  dst = &pd->dst
             .buf[(blk_row * pd->dst.stride + blk_col) << tx_size_wide_log2[0]];
605

606 607
  a = &args->ta[blk_col];
  l = &args->tl[blk_row];
608
  ctx = get_entropy_context(tx_size, a, l);
Jingning Han's avatar
Jingning Han committed
609

610
  // Assert not magic number (uninitialized).
611
  assert(x->blk_skip[plane][blk_row * bw + blk_col] != 234);
612

613
  if (x->blk_skip[plane][blk_row * bw + blk_col] == 0) {
Angie Chiang's avatar
Angie Chiang committed
614
    av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
615
                    ctx, AV1_XFORM_QUANT_FP);
616
  } else {
617
    p->eobs[block] = 0;
618
  }
619

Jingning Han's avatar
Jingning Han committed
620
  av1_optimize_b(cm, x, plane, blk_row, blk_col, block, plane_bsize, tx_size, a,
621
                 l, CONFIG_LV_MAP);
622

623 624
  av1_set_txb_context(x, plane, block, tx_size, a, l);

625
  if (p->eobs[block]) *(args->skip) = 0;
Jingning Han's avatar
Jingning Han committed
626

627
  if (p->eobs[block] != 0)
628

629 630 631 632
  {
    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
633
#if CONFIG_MRC_TX && SIGNAL_ANY_MRC_MASK
634
                                mrc_mask,
Sarah Parker's avatar
Sarah Parker committed
635
#endif  // CONFIG_MRC_TX && SIGNAL_ANY_MRC_MASK
636
                                plane, tx_type, tx_size, dst, pd->dst.stride,
637 638
                                p->eobs[block]);
  }
Jingning Han's avatar
Jingning Han committed
639 640
}

641 642 643 644 645 646 647
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;
648
  const BLOCK_SIZE bsize = txsize_to_bsize[tx_size];
649
  const struct macroblockd_plane *const pd = &xd->plane[plane];
650 651
  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
652
  TX_SIZE plane_tx_size;
653 654
  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
655

656
  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
657

658 659 660
  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
661

662
  if (tx_size == plane_tx_size) {
663
    encode_block(plane, block, blk_row, blk_col, plane_bsize, tx_size, arg);
664
  } else {
665
    assert(tx_size < TX_SIZES_ALL);
666
    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
667 668
    assert(IMPLIES(tx_size <= TX_4X4, sub_txs == tx_size));
    assert(IMPLIES(tx_size > TX_4X4, sub_txs < tx_size));
669
    // This is the square transform block partition entry point.
670 671 672 673
    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);
674

675 676 677 678
    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;
679

680
        if (offsetr >= max_blocks_high || offsetc >= max_blocks_wide) continue;
681

682 683 684 685
        encode_block_inter(plane, block, offsetr, offsetc, plane_bsize, sub_txs,
                           arg);
        block += step;
      }
686 687 688 689
    }
  }
}

Angie Chiang's avatar
Angie Chiang committed
690 691 692 693 694
typedef struct encode_block_pass1_args {
  AV1_COMMON *cm;
  MACROBLOCK *x;
} encode_block_pass1_args;

695
static void encode_block_pass1(int plane, int block, int blk_row, int blk_col,
696 697
                               BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
                               void *arg) {
Angie Chiang's avatar
Angie Chiang committed
698 699 700
  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
701 702 703 704
  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);
705
  TxfmParam txfm_param;
Jingning Han's avatar
Jingning Han committed
706
  uint8_t *dst;
707
  int ctx = 0;
708 709
  dst = &pd->dst
             .buf[(blk_row * pd->dst.stride + blk_col) << tx_size_wide_log2[0]];
Jingning Han's avatar
Jingning Han committed
710

Angie Chiang's avatar
Angie Chiang committed
711
  av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
712
                  ctx, AV1_XFORM_QUANT_B);
713

Yushin Cho's avatar
Yushin Cho committed
714
  if (p->eobs[block] > 0) {
715
    txfm_param.bd = xd->bd;
716
    txfm_param.is_hbd = get_bitdepth_data_path_index(xd);
717
    txfm_param.tx_type = DCT_DCT;
718
    txfm_param.tx_size = tx_size;
719 720
    txfm_param.eob = p->eobs[block];
    txfm_param.lossless = xd->lossless[xd->mi[0]->mbmi.segment_id];
721 722 723
    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);
724 725 726
#if CONFIG_DAALA_TX
    daala_inv_txfm_add(dqcoeff, dst, pd->dst.stride, &txfm_param);
#else
727
#if CONFIG_HIGHBITDEPTH
728
    if (txfm_param.is_hbd) {
729
      av1_highbd_inv_txfm_add_4x4(dqcoeff, dst, pd->dst.stride, &txfm_param);
730
      return;
Jingning Han's avatar
Jingning Han committed
731
    }
732
#endif  //  CONFIG_HIGHBITDEPTH
Yaowu Xu's avatar
Yaowu Xu committed
733
    if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {
734
      av1_iwht4x4_add(dqcoeff, dst, pd->dst.stride, &txfm_param);
735
    } else {
736
      av1_idct4x4_add(dqcoeff, dst, pd->dst.stride, &txfm_param);
737
    }
738
#endif
Jingning Han's avatar
Jingning Han committed
739 740 741
  }
}

Angie Chiang's avatar
Angie Chiang committed
742 743
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
744 745
  av1_subtract_plane(x, bsize, 0);
  av1_foreach_transformed_block_in_plane(&x->e_mbd, bsize, 0,
Angie Chiang's avatar
Angie Chiang committed
746
                                         encode_block_pass1, &args);
Jingning Han's avatar
Jingning Han committed
747 748
}

Yaowu Xu's avatar
Yaowu Xu committed
749 750
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
751 752 753
  MACROBLOCKD *const xd = &x->e_mbd;
  struct optimize_ctx ctx;
  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
Angie Chiang's avatar
Angie Chiang committed
754
  struct encode_b_args arg = { cm, x, &ctx, &mbmi->skip, NULL, NULL, 1 };
Jingning Han's avatar
Jingning Han committed
755 756 757 758
  int plane;

  mbmi->skip = 1;

759
  if (x->skip) return;
Jingning Han's avatar
Jingning Han committed
760 761

  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
762 763 764 765 766
    const int subsampling_x = xd->plane[plane].subsampling_x;
    const int subsampling_y = xd->plane[plane].subsampling_y;

    if (!is_chroma_reference(mi_row, mi_col, bsize, subsampling_x,
                             subsampling_y))
767
      continue;
768 769

    bsize = scale_chroma_bsize(bsize, subsampling_x, subsampling_y);