encodemb.c 34.1 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

Angie Chiang's avatar
Angie Chiang committed
21
22
23
24
#if CONFIG_BITSTREAM_DEBUG || CONFIG_MISMATCH_DEBUG
#include "aom_util/debug_util.h"
#endif  // CONFIG_BITSTREAM_DEBUG || CONFIG_MISMATCH_DEBUG

25
26
27
28
#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
29

30
#include "av1/encoder/av1_quantize.h"
31
#include "av1/encoder/encodemb.h"
32
33
34
#if CONFIG_LV_MAP
#include "av1/encoder/encodetxb.h"
#endif
35
#include "av1/encoder/hybrid_fwd_txfm.h"
36
37
38
#if CONFIG_DAALA_TX
#include "av1/common/daala_inv_txfm.h"
#endif
39
#include "av1/encoder/rd.h"
40
#include "av1/encoder/rdopt.h"
Jingning Han's avatar
Jingning Han committed
41

42
43
44
45
#if CONFIG_CFL
#include "av1/common/cfl.h"
#endif

46
47
48
// 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
49
50
51
52
53
54
55
56
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) {
  if (check_subtract_block_size(rows, cols)) {
    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);
57
58
      return;
    }
Angie Chiang's avatar
Angie Chiang committed
59
60
    aom_subtract_block_c(rows, cols, diff, diff_stride, src8, src_stride, pred8,
                         pred_stride);
61
62
63
64

    return;
  }

Angie Chiang's avatar
Angie Chiang committed
65
66
67
  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
68
69
    return;
  }
Angie Chiang's avatar
Angie Chiang committed
70
71
72
73
  aom_subtract_block(rows, cols, diff, diff_stride, src8, src_stride, pred8,
                     pred_stride);
}

Angie Chiang's avatar
Angie Chiang committed
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
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
94
95
96
97
98
99
100
101
102
103
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
104
105
}

106
#if !CONFIG_LV_MAP
107
108
109
110
111
112
113
114
115
116
117
// 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);
}

118
119
// These numbers are empirically obtained.
static const int plane_rd_mult[REF_TYPES][PLANE_TYPES] = {
Thomas Davies's avatar
Thomas Davies committed
120
  { 10, 7 }, { 8, 5 },
121
};
Jingning Han's avatar
Jingning Han committed
122

123
static int optimize_b_greedy(const AV1_COMMON *cm, MACROBLOCK *mb, int plane,
124
                             int blk_row, int blk_col, int block,
125
                             TX_SIZE tx_size, int ctx, int fast_mode) {
126
127
128
  MACROBLOCKD *const xd = &mb->e_mbd;
  struct macroblock_plane *const p = &mb->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
129
130
131
132
  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));
133
  assert(eob <= av1_get_max_eob(tx_size));
134
135
136
137
  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);
138
  const int16_t *const dequant_ptr = p->dequant_QTX;
139
  const uint8_t *const band_translate = get_band_translate(tx_size);
140
  const TX_TYPE tx_type =
Luc Trudeau's avatar
Luc Trudeau committed
141
      av1_get_tx_type(plane_type, xd, blk_row, blk_col, 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
  const int16_t *const scan = scan_order->scan;
  const int16_t *const nb = scan_order->neighbors;
146
147
148
149
150
151
#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
152
  const int shift = av1_get_tx_scale(tx_size);
153
#endif
154
155
#if CONFIG_AOM_QM
  int seg_id = xd->mi[0]->mbmi.segment_id;
156
  const TX_SIZE qm_tx_size = av1_get_adjusted_tx_size(tx_size);
157
  // Use a flat matrix (i.e. no weighting) for 1D and Identity transforms
158
159
160
161
162
  const qm_val_t *iqmatrix =
      IS_2D_TRANSFORM(tx_type)
          ? pd->seg_iqmatrix[seg_id][qm_tx_size]
          : cm->giqmatrix[NUM_QM_LEVELS - 1][0][qm_tx_size];
#endif  // CONFIG_AOM_QM
163
#if CONFIG_NEW_QUANT
164
  int dq = get_dq_profile(cm->dq_type, mb->qindex, ref, plane_type);
165
  const dequant_val_type_nuq *dequant_val = p->dequant_val_nuq_QTX[dq];
166
167
168
#endif  // CONFIG_NEW_QUANT
  int64_t rd_cost0, rd_cost1;
  int16_t t0, t1;
169
  int i, final_eob = 0;
170
  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, xd->bd);
hui su's avatar
hui su committed
171
  int(*head_token_costs)[COEFF_CONTEXTS][TAIL_TOKENS] =
172
      mb->token_head_costs[get_txsize_entropy_ctx(tx_size)][plane_type][ref];
hui su's avatar
hui su committed
173
  int(*tail_token_costs)[COEFF_CONTEXTS][TAIL_TOKENS] =
174
      mb->token_tail_costs[get_txsize_entropy_ctx(tx_size)][plane_type][ref];
Dake He's avatar
Dake He committed
175

176
  const int64_t rdmult = (mb->rdmult * plane_rd_mult[ref][plane_type]) >> 1;
177
  int64_t rate0, rate1;
178
179
180
181
182
  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];
183
184
185
186

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

187
188
  for (i = 0; i < eob; i++) {
    const int rc = scan[i];
189
    token_cache[rc] = av1_pt_energy_class[av1_get_token(qcoeff[rc])];
190
191
192
193
  }

  /* 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
  rate0 = head_token_costs[0][ctx][0];
198
  int64_t best_block_rd_cost = RDCOST(rdmult, rate0, accu_error);
199
200

  // int64_t best_block_rd_cost_all0 = best_block_rd_cost;
hui su's avatar
hui su committed
201
202
  const int seg_eob =
      av1_get_tx_eob(&cm->seg, xd->mi[0]->mbmi.segment_id, tx_size);
203
204
  for (i = 0; i < eob; i++) {
    const int rc = scan[i];
205
206
207
208
    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
209
    const int8_t eob_val =
hui su's avatar
hui su committed
210
211
        (i + 1 == eob) ? (i + 1 == seg_eob ? LAST_EOB : EARLY_EOB) : NO_EOB;
    const int is_first = (i == 0);
212
213
214

    if (x == 0) {
      // no need to search when x == 0
215
216
217
      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]);
218
219
      // accu_error does not change when x==0
    } else {
220
221
222
223
224
225
226
227
/*  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
228
      int dx0 = abs(coeff[rc]) * (1 << shift);
Sebastien Alaiwan's avatar
Sebastien Alaiwan committed
229
230
      dx0 >>= xd->bd - 8;

231
      const int64_t d0 = (int64_t)dx0 * dx0;
232
#endif
233
      const int x_a = x - 2 * sz - 1;
234
      int dqv = dequant_ptr[rc != 0];
235
#if CONFIG_AOM_QM
236
      if (iqmatrix != NULL) {
237
        const qm_val_t iwt = iqmatrix[rc];
238
239
        dqv = ((iwt * (int)dqv) + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
      }
240
#endif  // CONFIG_AOM_QM
241

242
243
244
245
#if CONFIG_DAALA_TX
      int dx = dqcoeff[rc] - coeff[rc];
      const int64_t d2 = ((int64_t)dx * dx + depth_round) >> depth_shift;
#else
246
      int dx = (dqcoeff[rc] - coeff[rc]) * (1 << shift);
247
      dx = signed_shift_right(dx, xd->bd - 8);
248
      const int64_t d2 = (int64_t)dx * dx;
249
#endif
250
251
252
253

      /* compute the distortion for the second candidate
       * x_a = x - 2 * sz + 1;
       */
254
      int64_t d2_a;
255
      if (x_a != 0) {
256
257
#if CONFIG_DAALA_TX
#if CONFIG_NEW_QUANT
258
259
        dx = av1_dequant_coeff_nuq(x_a, dqv, dequant_val[rc != 0], 0) -
             coeff[rc];
260
261
262
263
264
#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
265
#if CONFIG_NEW_QUANT
266
        dx = av1_dequant_coeff_nuq(x_a, dqv, dequant_val[rc != 0], 0) -
Yaowu Xu's avatar
Yaowu Xu committed
267
             (coeff[rc] * (1 << shift));
Sebastien Alaiwan's avatar
Sebastien Alaiwan committed
268
        dx >>= xd->bd - 8;
269
#else   // CONFIG_NEW_QUANT
Sebastien Alaiwan's avatar
Sebastien Alaiwan committed
270
        dx -= ((dqv >> (xd->bd - 8)) + sz) ^ sz;
271
272
#endif  // CONFIG_NEW_QUANT
        d2_a = (int64_t)dx * dx;
273
#endif  // CONFIG_DAALA_TX
274
275
276
      } else {
        d2_a = d0;
      }
277

278
      // Computing RD cost
hui su's avatar
hui su committed
279
      int64_t base_bits;
280
281
      // rate cost of x
      base_bits = av1_get_token_cost(x, &t0, cat6_bits);
hui su's avatar
hui su committed
282
283
284
285
      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]);
286
      // rate cost of x_a
287
      base_bits = av1_get_token_cost(x_a, &t1, cat6_bits);
hui su's avatar
hui su committed
288
289
290
291
292
293
294
295
      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]);
      }
296

hui su's avatar
hui su committed
297
      int64_t next_bits0 = 0, next_bits1 = 0;
298
      if (i < eob - 1) {
hui su's avatar
hui su committed
299
        int ctx_next;
300
301
        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
302
        const int8_t eob_val_next =
hui su's avatar
hui su committed
303
            (i + 2 == eob) ? (i + 2 == seg_eob ? LAST_EOB : EARLY_EOB) : NO_EOB;
304
305
306

        token_cache[rc] = av1_pt_energy_class[t0];
        ctx_next = get_coef_context(nb, token_cache, i + 1);
307
308
309
        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]);
310
311
312

        token_cache[rc] = av1_pt_energy_class[t1];
        ctx_next = get_coef_context(nb, token_cache, i + 1);
313
314
315
        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]);
316
317
      }

318
319
      rd_cost0 = RDCOST(rdmult, (rate0 + next_bits0), d2);
      rd_cost1 = RDCOST(rdmult, (rate1 + next_bits1), d2_a);
320
      const int best_x = (rd_cost1 < rd_cost0);
321

322
      const int eob_v = (i + 1 == seg_eob) ? LAST_EOB : EARLY_EOB;
hui su's avatar
hui su committed
323
324
325
326
327
328
329
      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));
330
331
      eob_cost1 = eob_cost0;
      if (x_a != 0) {
hui su's avatar
hui su committed
332
333
334
335
        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),
336
337
338
339
340
341
                           (accu_error + d2_a - d0));
        best_eob_x = (eob_cost1 < eob_cost0);
      } else {
        best_eob_x = 0;
      }

342
343
344
      const int dqc = dqcoeff[rc];
      int dqc_a = 0;
      if (best_x || best_eob_x) {
345
        if (x_a != 0) {
346
347
#if CONFIG_DAALA_TX
#if CONFIG_NEW_QUANT
348
349
          dqc_a =
              av1_dequant_abscoeff_nuq(abs(x_a), dqv, dequant_val[rc != 0], 0);
350
351
352
353
354
          if (sz) dqc_a = -dqc_a;
#else
          dqc_a = x_a * dqv;
#endif  // CONFIG_NEW_QUANT
#else   // CONFIG_DAALA_TX
355
#if CONFIG_NEW_QUANT
356
357
          dqc_a = av1_dequant_abscoeff_nuq(abs(x_a), dqv, dequant_val[rc != 0],
                                           shift);
358
359
          if (sz) dqc_a = -dqc_a;
#else
360
361
          if (x_a < 0)
            dqc_a = -((-x_a * dqv) >> shift);
362
          else
363
            dqc_a = (x_a * dqv) >> shift;
364
#endif  // CONFIG_NEW_QUANT
365
#endif  // CONFIG_DAALA_TX
366
367
        } else {
          dqc_a = 0;
368
        }
369
370
371
372
      }

      // record the better quantized value
      if (best_x) {
373
        assert(d2_a <= d0);
374
375
376
377
378
379
        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 {
380
        assert(d2 <= d0);
381
382
383
384
        accu_rate += rate0;
        accu_error += d2 - d0;
        token_cache[rc] = av1_pt_energy_class[t0];
      }
385
      assert(accu_error >= 0);
386
387

      // determine whether to move the eob position to i+1
388
389
      const int use_a = (x_a != 0) && (best_eob_x);
      const int64_t best_eob_cost_i = use_a ? eob_cost1 : eob_cost0;
390
391
392
      if (best_eob_cost_i < best_block_rd_cost) {
        best_block_rd_cost = best_eob_cost_i;
        final_eob = i + 1;
393
394
395
396
397
398
399
        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;
        }
400
401
402
403
404
405
      }
    }  // if (x==0)
  }    // for (i)

  assert(final_eob <= eob);
  if (final_eob > 0) {
406
    assert(before_best_eob_qc != 0);
407
408
    i = final_eob - 1;
    int rc = scan[i];
409
410
    qcoeff[rc] = before_best_eob_qc;
    dqcoeff[rc] = before_best_eob_dqc;
411
412
413
414
415
416
417
418
  }

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

419
  p->eobs[block] = final_eob;
420
421
  return final_eob;
}
422
#endif  // !CONFIG_LV_MAP
423

424
int av1_optimize_b(const struct AV1_COMP *cpi, MACROBLOCK *mb, int plane,
425
                   int blk_row, int blk_col, int block, BLOCK_SIZE plane_bsize,
Jingning Han's avatar
Jingning Han committed
426
                   TX_SIZE tx_size, const ENTROPY_CONTEXT *a,
427
                   const ENTROPY_CONTEXT *l, int fast_mode) {
428
429
430
  MACROBLOCKD *const xd = &mb->e_mbd;
  struct macroblock_plane *const p = &mb->plane[plane];
  const int eob = p->eobs[block];
431
432
  if (eob == 0 || !mb->optimize || xd->lossless[xd->mi[0]->mbmi.segment_id])
    return eob;
Jingning Han's avatar
Jingning Han committed
433

434
#if !CONFIG_LV_MAP
435
  (void)plane_bsize;
Jingning Han's avatar
Jingning Han committed
436
437
  (void)blk_row;
  (void)blk_col;
438
  int ctx = get_entropy_context(tx_size, a, l);
439
  const AV1_COMMON *const cm = &cpi->common;
440
441
  return optimize_b_greedy(cm, mb, plane, blk_row, blk_col, block, tx_size, ctx,
                           fast_mode);
442
443
444
#else   // !CONFIG_LV_MAP
  TXB_CTX txb_ctx;
  get_txb_ctx(plane_bsize, tx_size, plane, a, l, &txb_ctx);
445
  return av1_optimize_txb(cpi, mb, plane, blk_row, blk_col, block, tx_size,
446
                          &txb_ctx, fast_mode);
447
#endif  // !CONFIG_LV_MAP
448
}
449

Angie Chiang's avatar
Angie Chiang committed
450
451
typedef enum QUANT_FUNC {
  QUANT_FUNC_LOWBD = 0,
452
  QUANT_FUNC_HIGHBD = 1,
453
  QUANT_FUNC_TYPES = 2
Angie Chiang's avatar
Angie Chiang committed
454
455
} QUANT_FUNC;

456
457
static AV1_QUANT_FACADE
    quant_func_list[AV1_XFORM_QUANT_TYPES][QUANT_FUNC_TYPES] = {
458
#if !CONFIG_NEW_QUANT
459
      { av1_quantize_fp_facade, av1_highbd_quantize_fp_facade },
Yaowu Xu's avatar
Yaowu Xu committed
460
461
      { av1_quantize_b_facade, av1_highbd_quantize_b_facade },
      { av1_quantize_dc_facade, av1_highbd_quantize_dc_facade },
462
#else   // !CONFIG_NEW_QUANT
463
464
465
      { 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 },
466
#endif  // !CONFIG_NEW_QUANT
467
468
      { NULL, NULL }
    };
469

Angie Chiang's avatar
Angie Chiang committed
470
471
void av1_xform_quant(const AV1_COMMON *cm, MACROBLOCK *x, int plane, int block,
                     int blk_row, int blk_col, BLOCK_SIZE plane_bsize,
472
                     TX_SIZE tx_size, AV1_XFORM_QUANT xform_quant_idx) {
Jingning Han's avatar
Jingning Han committed
473
  MACROBLOCKD *const xd = &x->e_mbd;
474
  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
Jingning Han's avatar
Jingning Han committed
475
476
  const struct macroblock_plane *const p = &x->plane[plane];
  const struct macroblockd_plane *const pd = &xd->plane[plane];
477
  PLANE_TYPE plane_type = get_plane_type(plane);
Luc Trudeau's avatar
Luc Trudeau committed
478
  TX_TYPE tx_type = av1_get_tx_type(plane_type, xd, blk_row, blk_col, tx_size);
Angie Chiang's avatar
Angie Chiang committed
479

480
#if CONFIG_NEW_QUANT
481
  const int is_inter = is_inter_block(mbmi);
Angie Chiang's avatar
Angie Chiang committed
482
483
#endif
  const SCAN_ORDER *const scan_order = get_scan(cm, tx_size, tx_type, mbmi);
Jingning Han's avatar
Jingning Han committed
484
485
486
487
  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];
488
  const int diff_stride = block_size_wide[plane_bsize];
Yushin Cho's avatar
Yushin Cho committed
489
#if CONFIG_AOM_QM
490
  int seg_id = mbmi->segment_id;
491
  const TX_SIZE qm_tx_size = av1_get_adjusted_tx_size(tx_size);
492
  // Use a flat matrix (i.e. no weighting) for 1D and Identity transforms
493
494
495
496
497
498
499
  const qm_val_t *qmatrix =
      IS_2D_TRANSFORM(tx_type) ? pd->seg_qmatrix[seg_id][qm_tx_size]
                               : cm->gqmatrix[NUM_QM_LEVELS - 1][0][qm_tx_size];
  const qm_val_t *iqmatrix =
      IS_2D_TRANSFORM(tx_type)
          ? pd->seg_iqmatrix[seg_id][qm_tx_size]
          : cm->giqmatrix[NUM_QM_LEVELS - 1][0][qm_tx_size];
500
#endif  // CONFIG_AOM_QM
Angie Chiang's avatar
Angie Chiang committed
501

502
  TxfmParam txfm_param;
503
  QUANT_PARAM qparam;
504
505
  const int16_t *src_diff;

506
507
  src_diff =
      &p->src_diff[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
508
509
510
#if CONFIG_DAALA_TX
  qparam.log_scale = 0;
#else
511
  qparam.log_scale = av1_get_tx_scale(tx_size);
512
#endif
513
  qparam.tx_size = tx_size;
514
#if CONFIG_NEW_QUANT
515
  qparam.dq = get_dq_profile(cm->dq_type, x->qindex, is_inter, plane_type);
516
517
518
519
520
#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
521

522
523
524
  txfm_param.tx_type = tx_type;
  txfm_param.tx_size = tx_size;
  txfm_param.lossless = xd->lossless[mbmi->segment_id];
525
526
527
  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);
Angie Chiang's avatar
Angie Chiang committed
528

529
  txfm_param.bd = xd->bd;
530
  txfm_param.is_hbd = get_bitdepth_data_path_index(xd);
Angie Chiang's avatar
Angie Chiang committed
531

532
  av1_highbd_fwd_txfm(src_diff, coeff, diff_stride, &txfm_param);
533

Yaowu Xu's avatar
Yaowu Xu committed
534
  if (xform_quant_idx != AV1_XFORM_QUANT_SKIP_QUANT) {
535
    const int n_coeffs = av1_get_max_eob(tx_size);
536
    if (LIKELY(!x->skip_block)) {
537
#if CONFIG_DAALA_TX
538
      quant_func_list[xform_quant_idx][1](coeff, n_coeffs, p, qcoeff, dqcoeff,
539
540
                                          eob, scan_order, &qparam);
#else
541
      quant_func_list[xform_quant_idx][txfm_param.is_hbd](
542
          coeff, n_coeffs, p, qcoeff, dqcoeff, eob, scan_order, &qparam);
543
#endif
544
    } else {
545
      av1_quantize_skip(n_coeffs, qcoeff, dqcoeff, eob);
Jingning Han's avatar
Jingning Han committed
546
547
    }
  }
548
549
550
551
#if CONFIG_LV_MAP
  p->txb_entropy_ctx[block] =
      (uint8_t)av1_get_txb_entropy_context(qcoeff, scan_order, *eob);
#endif  // CONFIG_LV_MAP
552
  return;
Jingning Han's avatar
Jingning Han committed
553
554
}

555
static void encode_block(int plane, int block, int blk_row, int blk_col,
Angie Chiang's avatar
Angie Chiang committed
556
557
558
559
560
                         BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg,
                         int mi_row, int mi_col, RUN_TYPE dry_run) {
  (void)mi_row;
  (void)mi_col;
  (void)dry_run;
Jingning Han's avatar
Jingning Han committed
561
  struct encode_b_args *const args = arg;
562
  const AV1_COMMON *const cm = &args->cpi->common;
Jingning Han's avatar
Jingning Han committed
563
564
565
566
567
568
569
  MACROBLOCK *const x = args->x;
  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);
  uint8_t *dst;
  ENTROPY_CONTEXT *a, *l;
Yushin Cho's avatar
Yushin Cho committed
570

571
  int bw = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
572
573
  dst = &pd->dst
             .buf[(blk_row * pd->dst.stride + blk_col) << tx_size_wide_log2[0]];
574

575
576
  a = &args->ta[blk_col];
  l = &args->tl[blk_row];
Jingning Han's avatar
Jingning Han committed
577

578
  // Assert not magic number (uninitialized).
579
  assert(x->blk_skip[plane][blk_row * bw + blk_col] != 234);
580

581
  if (x->blk_skip[plane][blk_row * bw + blk_col] == 0) {
582
583
584
    if (args->enable_optimize_b) {
      av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize,
                      tx_size, AV1_XFORM_QUANT_FP);
585
586
      av1_optimize_b(args->cpi, x, plane, blk_row, blk_col, block, plane_bsize,
                     tx_size, a, l, CONFIG_LV_MAP);
587
588
589
590
    } else {
      av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize,
                      tx_size, AV1_XFORM_QUANT_B);
    }
591
  } else {
592
    p->eobs[block] = 0;
593
594
595
#if CONFIG_LV_MAP
    p->txb_entropy_ctx[block] = 0;
#endif
596
  }
597

598
599
  av1_set_txb_context(x, plane, block, tx_size, a, l);

600
601
  if (p->eobs[block]) {
    *(args->skip) = 0;
Jingning Han's avatar
Jingning Han committed
602

603
    TX_TYPE tx_type =
Luc Trudeau's avatar
Luc Trudeau committed
604
        av1_get_tx_type(pd->plane_type, xd, blk_row, blk_col, tx_size);
Sebastien Alaiwan's avatar
Sebastien Alaiwan committed
605
606
607
    av1_inverse_transform_block(xd, dqcoeff, plane, tx_type, tx_size, dst,
                                pd->dst.stride, p->eobs[block],
                                cm->reduced_tx_set_used);
608
  }
609
610
611

#if CONFIG_TXK_SEL
  if (plane == 0 && p->eobs[block] == 0) {
612
613
614
615
616
617
618
619
    if (args->enable_optimize_b) {
      xd->mi[0]->mbmi.txk_type[(blk_row << MAX_MIB_SIZE_LOG2) + blk_col] =
          DCT_DCT;
    } else {
      assert(
          xd->mi[0]->mbmi.txk_type[(blk_row << MAX_MIB_SIZE_LOG2) + blk_col] ==
          DCT_DCT);
    }
620
621
622
  }
#endif  // CONFIG_TXK_SEL

Angie Chiang's avatar
Angie Chiang committed
623
624
625
#if CONFIG_MISMATCH_DEBUG
  if (dry_run == OUTPUT_ENABLED) {
    int pixel_c, pixel_r;
626
627
628
    BLOCK_SIZE bsize = txsize_to_bsize[tx_size];
    int blk_w = block_size_wide[bsize];
    int blk_h = block_size_high[bsize];
Angie Chiang's avatar
Angie Chiang committed
629
630
631
632
633
634
    mi_to_pixel_loc(&pixel_c, &pixel_r, mi_col, mi_row, blk_col, blk_row,
                    pd->subsampling_x, pd->subsampling_y);
    mismatch_record_block_tx(dst, pd->dst.stride, plane, pixel_c, pixel_r,
                             blk_w, blk_h);
  }
#endif
Jingning Han's avatar
Jingning Han committed
635
636
}

637
638
static void encode_block_inter(int plane, int block, int blk_row, int blk_col,
                               BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
Angie Chiang's avatar
Angie Chiang committed
639
640
641
642
                               void *arg, int mi_row, int mi_col,
                               RUN_TYPE dry_run) {
  (void)mi_row;
  (void)mi_col;
643
644
645
646
647
  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;
  const struct macroblockd_plane *const pd = &xd->plane[plane];
648
649
  const int tx_row = blk_row >> (1 - pd->subsampling_y);
  const int tx_col = blk_col >> (1 - pd->subsampling_x);
650
651
  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
652

653
  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
654

655
  const TX_SIZE plane_tx_size =
656
      plane ? av1_get_uv_tx_size(mbmi, pd->subsampling_x, pd->subsampling_y)
657
            : mbmi->inter_tx_size[tx_row][tx_col];
Debargha Mukherjee's avatar
Debargha Mukherjee committed
658

659
  if (tx_size == plane_tx_size || plane) {
Angie Chiang's avatar
Angie Chiang committed
660
661
    encode_block(plane, block, blk_row, blk_col, plane_bsize, tx_size, arg,
                 mi_row, mi_col, dry_run);
662
  } else {
663
    assert(tx_size < TX_SIZES_ALL);
664
    const TX_SIZE sub_txs = sub_tx_size_map[1][tx_size];
665
666
    assert(IMPLIES(tx_size <= TX_4X4, sub_txs == tx_size));
    assert(IMPLIES(tx_size > TX_4X4, sub_txs < tx_size));
667
    // This is the square transform block partition entry point.
668
669
670
671
    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);
672

673
674
675
676
    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;
677

678
        if (offsetr >= max_blocks_high || offsetc >= max_blocks_wide) continue;
679

680
        encode_block_inter(plane, block, offsetr, offsetc, plane_bsize, sub_txs,
Angie Chiang's avatar
Angie Chiang committed
681
                           arg, mi_row, mi_col, dry_run);
682
683
        block += step;
      }
684
685
686
687
    }
  }
}

Angie Chiang's avatar
Angie Chiang committed
688
689
690
691
692
typedef struct encode_block_pass1_args {
  AV1_COMMON *cm;
  MACROBLOCK *x;
} encode_block_pass1_args;

693
static void encode_block_pass1(int plane, int block, int blk_row, int blk_col,
694
695
                               BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
                               void *arg) {
Angie Chiang's avatar
Angie Chiang committed
696
697
698
  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
699
700
701
702
  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);
703
  TxfmParam txfm_param;
Jingning Han's avatar
Jingning Han committed
704
  uint8_t *dst;
705
706
  dst = &pd->dst
             .buf[(blk_row * pd->dst.stride + blk_col) << tx_size_wide_log2[0]];
Jingning Han's avatar
Jingning Han committed
707

Angie Chiang's avatar
Angie Chiang committed
708
  av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
709
                  AV1_XFORM_QUANT_B);
710

Yushin Cho's avatar
Yushin Cho committed
711
  if (p->eobs[block] > 0) {
712
    txfm_param.bd = xd->bd;
713
    txfm_param.is_hbd = get_bitdepth_data_path_index(xd);
714
    txfm_param.tx_type = DCT_DCT;
715
    txfm_param.tx_size = tx_size;
716
717
    txfm_param.eob = p->eobs[block];
    txfm_param.lossless = xd->lossless[xd->mi[0]->mbmi.segment_id];
718
719
720
    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);
721
722
723
#if CONFIG_DAALA_TX
    daala_inv_txfm_add(dqcoeff, dst, pd->dst.stride, &txfm_param);
#else
724
    if (txfm_param.is_hbd) {
725
      av1_highbd_inv_txfm_add_4x4(dqcoeff, dst, pd->dst.stride, &txfm_param);
726
      return;
Jingning Han's avatar
Jingning Han committed
727
    }
Yaowu Xu's avatar
Yaowu Xu committed
728
    if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {
729
      av1_iwht4x4_add(dqcoeff, dst, pd->dst.stride, &txfm_param);
730
    } else {
731
      av1_idct4x4_add(dqcoeff, dst, pd->dst.stride, &txfm_param);
732
    }
733
#endif
Jingning Han's avatar
Jingning Han committed
734
735
736
  }
}

Angie Chiang's avatar
Angie Chiang committed
737
738
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
739
740
  av1_subtract_plane(x, bsize, 0);
  av1_foreach_transformed_block_in_plane(&x->e_mbd, bsize, 0,
Angie Chiang's avatar
Angie Chiang committed
741
                                         encode_block_pass1, &args);
Jingning Han's avatar
Jingning Han committed
742
743
}

744
void av1_encode_sb(const struct AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
745
                   int mi_row, int mi_col, RUN_TYPE dry_run) {
Angie Chiang's avatar
Angie Chiang committed
746
  (void)dry_run;
Jingning Han's avatar
Jingning Han committed
747
748
749
  MACROBLOCKD *const xd = &x->e_mbd;
  struct optimize_ctx ctx;
  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
750
751
  struct encode_b_args arg = { cpi,  x,    &ctx,       &mbmi->skip,
                               NULL, NULL, x->optimize };
Jingning Han's avatar
Jingning Han committed
752
753
754
755
  int plane;

  mbmi->skip = 1;

756
  if (x->skip) return;
Jingning Han's avatar
Jingning Han committed
757
758

  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
759
760
761
762
763
    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))
764
      continue;
765

766
767
    const BLOCK_SIZE bsizec =
        scale_chroma_bsize(bsize, subsampling_x, subsampling_y);
768

769
770
    // TODO(jingning): Clean this up.
    const struct macroblockd_plane *const pd = &xd->plane[plane];
771
    const BLOCK_SIZE plane_bsize = get_plane_block_size(bsizec, pd);
772
773
    const int mi_width = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
    const int mi_height = block_size_high[plane_bsize] >> tx_size_wide_log2[0];
774
    TX_SIZE max_tx_size = get_vartx_max_txsize(
775
        xd, plane_bsize, pd->subsampling_x || pd->subsampling_y);
776

777
    const BLOCK_SIZE txb_size = txsize_to_bsize[max_tx_size];
778
779
    const int bw = block_size_wide[txb_size] >> tx_size_wide_log2[0];
    const int bh = block_size_high[txb_size] >> tx_size_wide_log2[0];
780
781
    int idx, idy;
    int block = 0;
782
    int step = tx_size_wide_unit[max_tx_size] * tx_size_high_unit[max_tx_size];
783
    av1_get_entropy_contexts(bsizec, 0, pd, ctx.ta[plane], ctx.tl[plane]);
784

785
    av1_subtract_plane(x, bsizec, plane);
Yushin Cho's avatar
Yushin Cho committed
786

787
788
    arg.ta = ctx.ta[plane];
    arg.tl = ctx.tl[plane];
789

790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
    const BLOCK_SIZE max_unit_bsize = get_plane_block_size(BLOCK_64X64, pd);
    int mu_blocks_wide =
        block_size_wide[max_unit_bsize] >> tx_size_wide_log2[0];
    int mu_blocks_high =
        block_size_high[max_unit_bsize] >> tx_size_high_log2[0];

    mu_blocks_wide = AOMMIN(mi_width, mu_blocks_wide);
    mu_blocks_high = AOMMIN(mi_height, mu_blocks_high);

    for (idy = 0; idy < mi_height; idy += mu_blocks_high) {
      for (idx = 0; idx < mi_width; idx += mu_blocks_wide) {
        int blk_row, blk_col;
        const int unit_height = AOMMIN(mu_blocks_high + idy, mi_height);
        const int unit_width = AOMMIN(mu_blocks_wide + idx, mi_width);
        for (blk_row = idy; blk_row < unit_height; blk_row += bh) {
          for (blk_col = idx; blk_col < unit_width; blk_col += bw) {
            encode_block_inter(plane, block, blk_row, blk_col, plane_bsize,
Angie Chiang's avatar
Angie Chiang committed
807
                               max_tx_size, &arg, mi_row, mi_col, dry_run);
808
809
810
            block += step;
          }
        }
811
812
      }
    }
Jingning Han's avatar
Jingning Han committed
813
814
815
  }
}

816
void av1_set_txb_context(MACROBLOCK *x, int plane, int block, TX_SIZE tx_size,
817
818
819
                         ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l) {
  (void)tx_size;
  struct macroblock_plane *p = &x->plane[plane];
820

821
#if !CONFIG_LV_MAP
822
  *a = *l = p->eobs[block] > 0;
823
824
825
#else   // !CONFIG_LV_MAP
  *a = *l = p->txb_entropy_ctx[block];
#endif  // !CONFIG_LV_MAP
826
827
828
829
830

  int i;
  for (i = 0; i < tx_size_wide_unit[tx_size]; ++i) a[i] = a[0];

  for (i = 0; i < tx_size_high_unit[tx_size]; ++i) l[i] = l[0];
831
832
833
834
835
836
837
838
}

static void encode_block_intra_and_set_context(int plane, int block,
                                               int blk_row, int blk_col,
                                               BLOCK_SIZE plane_bsize,
                                               TX_SIZE tx_size, void *arg) {
  av1_encode_block_intra(plane, block, blk_row, blk_col, plane_bsize, tx_size,
                         arg);
Yushin Cho's avatar
Yushin Cho committed
839

840
841
842
843
  struct encode_b_args *const args = arg;
  MACROBLOCK *x = args->x;
  ENTROPY_CONTEXT *a = &args->ta[blk_col];
  ENTROPY_CONTEXT *l = &args->tl[blk_row];
844
  av1_set_txb_context(x, plane, block, tx_size, a, l);
845
846
}

Yaowu Xu's avatar
Yaowu Xu committed
847
848
849
void av1_encode_block_intra(int plane, int block, int blk_row, int blk_col,
                            BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
                            void *arg) {
850
  struct encode_b_args *const args = arg;
851
  const AV1_COMMON *const cm = &args->cpi->common;
Jingning Han's avatar
Jingning Han committed
852
853
854
855
856
  MACROBLOCK *const x = args->x;
  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 *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
857
  PLANE_TYPE plane_type = get_plane_type(plane);
858
  const TX_TYPE tx_type =
Luc Trudeau's avatar
Luc Trudeau committed
859
      av1_get_tx_type(plane_type, xd, blk_row, blk_col, tx_size);
Jingning Han's avatar
Jingning Han committed
860
861
  uint16_t *eob = &p->eobs[block];
  const int dst_stride = pd->dst.stride;
Angie Chiang's avatar
Angie Chiang committed
862
863
  uint8_t *dst =
      &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
864

Luc Trudeau's avatar
Luc Trudeau committed
865
  av1_predict_intra_block_facade(cm, xd, plane, blk_col, blk_row, tx_size);