encodemb.c 44.2 KB
Newer Older
Jingning Han's avatar
Jingning Han committed
1
/*
Yaowu Xu's avatar
Yaowu Xu committed
2
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
Jingning Han's avatar
Jingning Han committed
3
 *
Yaowu Xu's avatar
Yaowu Xu committed
4 5 6 7 8 9
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
Jingning Han's avatar
Jingning Han committed
10 11
 */

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

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

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

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

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

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

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

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

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

    return;
  }

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

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

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

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

113
#if !CONFIG_LV_MAP
114 115
// These numbers are empirically obtained.
static const int plane_rd_mult[REF_TYPES][PLANE_TYPES] = {
Thomas Davies's avatar
Thomas Davies committed
116
  { 10, 7 }, { 8, 5 },
117
};
Jingning Han's avatar
Jingning Han committed
118

119
static int optimize_b_greedy(const AV1_COMMON *cm, MACROBLOCK *mb, int plane,
120 121
                             int blk_row, int blk_col, int block,
                             TX_SIZE tx_size, int ctx) {
122 123 124
  MACROBLOCKD *const xd = &mb->e_mbd;
  struct macroblock_plane *const p = &mb->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
125 126 127 128 129
  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]);
130 131 132 133 134 135
  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);
  const int16_t *const dequant_ptr = pd->dequant;
  const uint8_t *const band_translate = get_band_translate(tx_size);
136
  const TX_TYPE tx_type =
137
      av1_get_tx_type(plane_type, xd, blk_row, blk_col, block, tx_size);
138
  const SCAN_ORDER *const scan_order =
Angie Chiang's avatar
Angie Chiang committed
139
      get_scan(cm, tx_size, tx_type, &xd->mi[0]->mbmi);
140 141
  const int16_t *const scan = scan_order->scan;
  const int16_t *const nb = scan_order->neighbors;
142
  const int shift = av1_get_tx_scale(tx_size);
143 144
#if CONFIG_AOM_QM
  int seg_id = xd->mi[0]->mbmi.segment_id;
145 146 147 148 149
  // Use a flat matrix (i.e. no weighting) for 1D and Identity transforms
  const qm_val_t *iqmatrix =
      IS_2D_TRANSFORM(tx_type)
          ? pd->seg_iqmatrix[seg_id][!ref][tx_size]
          : cm->giqmatrix[NUM_QM_LEVELS - 1][0][0][tx_size];
150 151 152 153 154 155 156
#endif
#if CONFIG_NEW_QUANT
  int dq = get_dq_profile_from_ctx(mb->qindex, ctx, ref, plane_type);
  const dequant_val_type_nuq *dequant_val = pd->dequant_val_nuq[dq];
#endif  // CONFIG_NEW_QUANT
  int64_t rd_cost0, rd_cost1;
  int16_t t0, t1;
157
  int i, final_eob = 0;
158
  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, xd->bd);
hui su's avatar
hui su committed
159 160 161 162
  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];
163
  const int64_t rdmult = (mb->rdmult * plane_rd_mult[ref][plane_type]) >> 1;
164
  int64_t rate0, rate1;
165 166 167 168 169
  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];
170 171
  for (i = 0; i < eob; i++) {
    const int rc = scan[i];
172
    token_cache[rc] = av1_pt_energy_class[av1_get_token(qcoeff[rc])];
173 174 175 176
  }

  /* Record the r-d cost */
  int64_t accu_rate = 0;
177 178 179
  // 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;
180
  rate0 = head_token_costs[0][ctx][0];
181
  int64_t best_block_rd_cost = RDCOST(rdmult, rate0, accu_error);
182 183

  // int64_t best_block_rd_cost_all0 = best_block_rd_cost;
hui su's avatar
hui su committed
184 185
  const int seg_eob =
      av1_get_tx_eob(&cm->seg, xd->mi[0]->mbmi.segment_id, tx_size);
186 187
  for (i = 0; i < eob; i++) {
    const int rc = scan[i];
188 189 190 191
    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);
hui su's avatar
hui su committed
192 193 194
    const int eob_val =
        (i + 1 == eob) ? (i + 1 == seg_eob ? LAST_EOB : EARLY_EOB) : NO_EOB;
    const int is_first = (i == 0);
195 196 197

    if (x == 0) {
      // no need to search when x == 0
198 199 200
      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]);
201 202 203 204 205 206
      // 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.
207
      int dx0 = abs(coeff[rc]) * (1 << shift);
Sebastien Alaiwan's avatar
Sebastien Alaiwan committed
208 209
      dx0 >>= xd->bd - 8;

210 211 212
      const int64_t d0 = (int64_t)dx0 * dx0;
      const int x_a = x - 2 * sz - 1;
      int dqv;
213
#if CONFIG_AOM_QM
214
      int iwt;
215
      dqv = dequant_ptr[rc != 0];
216 217 218 219
      if (iqmatrix != NULL) {
        iwt = iqmatrix[rc];
        dqv = ((iwt * (int)dqv) + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
      }
220 221 222 223
#else
      dqv = dequant_ptr[rc != 0];
#endif

224
      int dx = (dqcoeff[rc] - coeff[rc]) * (1 << shift);
Sebastien Alaiwan's avatar
Sebastien Alaiwan committed
225
      dx >>= xd->bd - 8;
226
      const int64_t d2 = (int64_t)dx * dx;
227 228 229 230

      /* compute the distortion for the second candidate
       * x_a = x - 2 * sz + 1;
       */
231
      int64_t d2_a;
232 233 234
      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
235
             (coeff[rc] * (1 << shift));
Sebastien Alaiwan's avatar
Sebastien Alaiwan committed
236
        dx >>= xd->bd - 8;
237
#else   // CONFIG_NEW_QUANT
Sebastien Alaiwan's avatar
Sebastien Alaiwan committed
238
        dx -= ((dqv >> (xd->bd - 8)) + sz) ^ sz;
239 240 241 242 243
#endif  // CONFIG_NEW_QUANT
        d2_a = (int64_t)dx * dx;
      } else {
        d2_a = d0;
      }
244
      // Computing RD cost
hui su's avatar
hui su committed
245
      int64_t base_bits;
246 247
      // rate cost of x
      base_bits = av1_get_token_cost(x, &t0, cat6_bits);
hui su's avatar
hui su committed
248 249 250 251
      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]);
252
      // rate cost of x_a
253
      base_bits = av1_get_token_cost(x_a, &t1, cat6_bits);
hui su's avatar
hui su committed
254 255 256 257 258 259 260 261
      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]);
      }
262

hui su's avatar
hui su committed
263
      int64_t next_bits0 = 0, next_bits1 = 0;
264
      if (i < eob - 1) {
hui su's avatar
hui su committed
265
        int ctx_next;
266 267
        const int band_next = band_translate[i + 1];
        const int token_next = av1_get_token(qcoeff[scan[i + 1]]);
hui su's avatar
hui su committed
268 269
        const int eob_val_next =
            (i + 2 == eob) ? (i + 2 == seg_eob ? LAST_EOB : EARLY_EOB) : NO_EOB;
270 271 272

        token_cache[rc] = av1_pt_energy_class[t0];
        ctx_next = get_coef_context(nb, token_cache, i + 1);
273 274 275
        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]);
276 277 278

        token_cache[rc] = av1_pt_energy_class[t1];
        ctx_next = get_coef_context(nb, token_cache, i + 1);
279 280 281
        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]);
282 283
      }

284 285
      rd_cost0 = RDCOST(rdmult, (rate0 + next_bits0), d2);
      rd_cost1 = RDCOST(rdmult, (rate1 + next_bits1), d2_a);
286
      const int best_x = (rd_cost1 < rd_cost0);
287

288
      const int eob_v = (i + 1 == seg_eob) ? LAST_EOB : EARLY_EOB;
hui su's avatar
hui su committed
289 290 291 292 293 294 295
      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));
296 297
      eob_cost1 = eob_cost0;
      if (x_a != 0) {
hui su's avatar
hui su committed
298 299 300 301
        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),
302 303 304 305 306 307
                           (accu_error + d2_a - d0));
        best_eob_x = (eob_cost1 < eob_cost0);
      } else {
        best_eob_x = 0;
      }

308 309 310
      const int dqc = dqcoeff[rc];
      int dqc_a = 0;
      if (best_x || best_eob_x) {
311 312 313 314 315 316 317
        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
318 319
          if (x_a < 0)
            dqc_a = -((-x_a * dqv) >> shift);
320
          else
321
            dqc_a = (x_a * dqv) >> shift;
322 323 324 325 326 327 328 329
#endif  // CONFIG_NEW_QUANT
        } else {
          dqc_a = 0;
        }  // if (x_a != 0)
      }

      // record the better quantized value
      if (best_x) {
330
        assert(d2_a <= d0);
331 332 333 334 335 336
        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 {
337
        assert(d2 <= d0);
338 339 340 341
        accu_rate += rate0;
        accu_error += d2 - d0;
        token_cache[rc] = av1_pt_energy_class[t0];
      }
342
      assert(accu_error >= 0);
343 344

      // determine whether to move the eob position to i+1
345 346
      const int use_a = (x_a != 0) && (best_eob_x);
      const int64_t best_eob_cost_i = use_a ? eob_cost1 : eob_cost0;
347 348 349
      if (best_eob_cost_i < best_block_rd_cost) {
        best_block_rd_cost = best_eob_cost_i;
        final_eob = i + 1;
350 351 352 353 354 355 356
        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;
        }
357 358 359 360 361 362
      }
    }  // if (x==0)
  }    // for (i)

  assert(final_eob <= eob);
  if (final_eob > 0) {
363
    assert(before_best_eob_qc != 0);
364 365
    i = final_eob - 1;
    int rc = scan[i];
366 367
    qcoeff[rc] = before_best_eob_qc;
    dqcoeff[rc] = before_best_eob_dqc;
368 369 370 371 372 373 374 375
  }

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

376
  p->eobs[block] = final_eob;
377 378
  return final_eob;
}
379
#endif  // !CONFIG_LV_MAP
380

Jingning Han's avatar
Jingning Han committed
381 382 383
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,
384
                   const ENTROPY_CONTEXT *l, int fast_mode) {
385 386 387 388 389 390
  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
391

392
#if CONFIG_PVQ
393 394
  (void)cm;
  (void)tx_size;
395 396
  (void)a;
  (void)l;
397 398
  return eob;
#endif
Jingning Han's avatar
Jingning Han committed
399

400
#if !CONFIG_LV_MAP
401
  (void)plane_bsize;
Jingning Han's avatar
Jingning Han committed
402 403
  (void)blk_row;
  (void)blk_col;
404
  (void)fast_mode;
405 406 407 408
#if CONFIG_VAR_TX
  int ctx = get_entropy_context(tx_size, a, l);
#else
  int ctx = combine_entropy_contexts(*a, *l);
409
#endif  // CONFIG_VAR_TX
410 411
  return optimize_b_greedy(cm, mb, plane, blk_row, blk_col, block, tx_size,
                           ctx);
412 413 414
#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
415
  return av1_optimize_txb(cm, mb, plane, blk_row, blk_col, block, tx_size,
416
                          &txb_ctx, fast_mode);
417
#endif  // !CONFIG_LV_MAP
418
}
419

Thomas Daede's avatar
Thomas Daede committed
420
#if !CONFIG_PVQ
Angie Chiang's avatar
Angie Chiang committed
421 422
typedef enum QUANT_FUNC {
  QUANT_FUNC_LOWBD = 0,
423
  QUANT_FUNC_HIGHBD = 1,
424
  QUANT_FUNC_TYPES = 2
Angie Chiang's avatar
Angie Chiang committed
425 426
} QUANT_FUNC;

427 428
static AV1_QUANT_FACADE
    quant_func_list[AV1_XFORM_QUANT_TYPES][QUANT_FUNC_TYPES] = {
429
#if !CONFIG_NEW_QUANT
430
      { av1_quantize_fp_facade, av1_highbd_quantize_fp_facade },
Yaowu Xu's avatar
Yaowu Xu committed
431 432
      { av1_quantize_b_facade, av1_highbd_quantize_b_facade },
      { av1_quantize_dc_facade, av1_highbd_quantize_dc_facade },
433
#else   // !CONFIG_NEW_QUANT
434 435 436
      { 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 },
437
#endif  // !CONFIG_NEW_QUANT
438 439
      { NULL, NULL }
    };
440
#endif  // !CONFIG_PVQ
441

442
#if !CONFIG_TXMG && !CONFIG_PVQ
443
typedef void (*fwdTxfmFunc)(const int16_t *diff, tran_low_t *coeff, int stride,
444
                            TxfmParam *txfm_param);
445 446
static const fwdTxfmFunc fwd_txfm_func[2] = { av1_fwd_txfm,
                                              av1_highbd_fwd_txfm };
Angie Chiang's avatar
Angie Chiang committed
447
#endif
448

Angie Chiang's avatar
Angie Chiang committed
449 450
void av1_xform_quant(const AV1_COMMON *cm, MACROBLOCK *x, int plane, int block,
                     int blk_row, int blk_col, BLOCK_SIZE plane_bsize,
451 452
                     TX_SIZE tx_size, int ctx,
                     AV1_XFORM_QUANT xform_quant_idx) {
Jingning Han's avatar
Jingning Han committed
453
  MACROBLOCKD *const xd = &x->e_mbd;
454
  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
Yushin Cho's avatar
Yushin Cho committed
455
#if !(CONFIG_PVQ || CONFIG_DIST_8X8)
Jingning Han's avatar
Jingning Han committed
456 457
  const struct macroblock_plane *const p = &x->plane[plane];
  const struct macroblockd_plane *const pd = &xd->plane[plane];
458 459 460 461
#else
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
#endif
462
  PLANE_TYPE plane_type = get_plane_type(plane);
463 464
  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
465

Thomas Davies's avatar
Thomas Davies committed
466
#if (CONFIG_AOM_QM || CONFIG_NEW_QUANT) && !CONFIG_PVQ
467
  const int is_inter = is_inter_block(mbmi);
Angie Chiang's avatar
Angie Chiang committed
468 469 470
#endif

  const SCAN_ORDER *const scan_order = get_scan(cm, tx_size, tx_type, mbmi);
Jingning Han's avatar
Jingning Han committed
471 472 473 474
  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];
475
  const int diff_stride = block_size_wide[plane_bsize];
Thomas Davies's avatar
Thomas Davies committed
476
#if CONFIG_AOM_QM && !CONFIG_PVQ
477
  int seg_id = mbmi->segment_id;
478 479 480 481 482 483 484 485
  // Use a flat matrix (i.e. no weighting) for 1D and Identity transforms
  const qm_val_t *qmatrix =
      IS_2D_TRANSFORM(tx_type) ? pd->seg_qmatrix[seg_id][!is_inter][tx_size]
                               : cm->gqmatrix[NUM_QM_LEVELS - 1][0][0][tx_size];
  const qm_val_t *iqmatrix =
      IS_2D_TRANSFORM(tx_type)
          ? pd->seg_iqmatrix[seg_id][!is_inter][tx_size]
          : cm->giqmatrix[NUM_QM_LEVELS - 1][0][0][tx_size];
486
#endif
Angie Chiang's avatar
Angie Chiang committed
487

488
  TxfmParam txfm_param;
489

490
#if CONFIG_PVQ || CONFIG_DIST_8X8 || CONFIG_LGT || CONFIG_MRC_TX
Yushin Cho's avatar
Yushin Cho committed
491 492
  uint8_t *dst;
  const int dst_stride = pd->dst.stride;
Yushin Cho's avatar
Yushin Cho committed
493
#if CONFIG_PVQ || CONFIG_DIST_8X8
Lester Lu's avatar
Lester Lu committed
494
  int16_t *pred;
495 496
  const int txw = tx_size_wide[tx_size];
  const int txh = tx_size_high[tx_size];
Yushin Cho's avatar
Yushin Cho committed
497 498
  int i, j;
#endif
Lester Lu's avatar
Lester Lu committed
499
#endif
Yushin Cho's avatar
Yushin Cho committed
500

501 502
#if !CONFIG_PVQ
  const int tx2d_size = tx_size_2d[tx_size];
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
  qparam.log_scale = av1_get_tx_scale(tx_size);
509 510 511 512 513 514 515 516
#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
517 518 519 520
#else
  tran_low_t *ref_coeff = BLOCK_OFFSET(pd->pvq_ref_coeff, block);
  int skip = 1;
  PVQ_INFO *pvq_info = NULL;
Yushin Cho's avatar
Yushin Cho committed
521 522 523
  uint8_t *src;
  int16_t *src_int16;
  const int src_stride = p->src.stride;
524

525
  (void)ctx;
526 527 528 529 530 531 532
  (void)scan_order;
  (void)qcoeff;

  if (x->pvq_coded) {
    assert(block < MAX_PVQ_BLOCKS_IN_SB);
    pvq_info = &x->pvq[block][plane];
  }
533 534 535
  src = &p->src.buf[(blk_row * src_stride + blk_col) << tx_size_wide_log2[0]];
  src_int16 =
      &p->src_int16[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
Yushin Cho's avatar
Yushin Cho committed
536

537
#if CONFIG_HIGHBITDEPTH
Thomas Daede's avatar
Thomas Daede committed
538
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
539 540
    for (j = 0; j < txh; j++)
      for (i = 0; i < txw; i++)
Thomas Daede's avatar
Thomas Daede committed
541 542 543
        src_int16[diff_stride * j + i] =
            CONVERT_TO_SHORTPTR(src)[src_stride * j + i];
  } else {
544
#endif  // CONFIG_HIGHBITDEPTH
545 546
    for (j = 0; j < txh; j++)
      for (i = 0; i < txw; i++)
Thomas Daede's avatar
Thomas Daede committed
547
        src_int16[diff_stride * j + i] = src[src_stride * j + i];
548
#if CONFIG_HIGHBITDEPTH
Thomas Daede's avatar
Thomas Daede committed
549
  }
550
#endif  // CONFIG_HIGHBITDEPTH
Yushin Cho's avatar
Yushin Cho committed
551 552
#endif

553
#if CONFIG_PVQ || CONFIG_DIST_8X8 || CONFIG_LGT || CONFIG_MRC_TX
Yushin Cho's avatar
Yushin Cho committed
554
  dst = &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
555 556
#endif  // CONFIG_PVQ || CONFIG_DIST_8X8 || CONFIG_LGT || CONFIG_MRC_TX

Yushin Cho's avatar
Yushin Cho committed
557
#if CONFIG_PVQ || CONFIG_DIST_8X8
558 559 560 561 562 563
  if (CONFIG_PVQ
#if CONFIG_DIST_8X8
      || x->using_dist_8x8
#endif  // CONFIG_DIST_8X8
      ) {
    pred = &pd->pred[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
564

Thomas Daede's avatar
Thomas Daede committed
565 566
// copy uint8 orig and predicted block to int16 buffer
// in order to use existing VP10 transform functions
567
#if CONFIG_HIGHBITDEPTH
568 569 570 571 572 573
    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 {
574
#endif  // CONFIG_HIGHBITDEPTH
575 576 577
      for (j = 0; j < txh; j++)
        for (i = 0; i < txw; i++)
          pred[diff_stride * j + i] = dst[dst_stride * j + i];
578
#if CONFIG_HIGHBITDEPTH
579
    }
580
#endif  // CONFIG_HIGHBITDEPTH
581
  }
Yushin Cho's avatar
Yushin Cho committed
582
#endif  // CONFIG_PVQ || CONFIG_DIST_8X8
Yushin Cho's avatar
Yushin Cho committed
583

584
  (void)ctx;
585

586 587 588
  txfm_param.tx_type = tx_type;
  txfm_param.tx_size = tx_size;
  txfm_param.lossless = xd->lossless[mbmi->segment_id];
589
#if CONFIG_MRC_TX || CONFIG_LGT
590
  txfm_param.is_inter = is_inter_block(mbmi);
591 592
  txfm_param.dst = dst;
  txfm_param.stride = dst_stride;
593 594
#if CONFIG_MRC_TX
  txfm_param.valid_mask = &mbmi->valid_mrc_mask;
Sarah Parker's avatar
Sarah Parker committed
595 596 597
#if SIGNAL_ANY_MRC_MASK
  txfm_param.mask = BLOCK_OFFSET(xd->mrc_mask, block);
#endif  // SIGNAL_ANY_MRC_MASK
598
#endif  // CONFIG_MRC_TX
599 600
#endif  // CONFIG_MRC_TX || CONFIG_LGT
#if CONFIG_LGT
Lester Lu's avatar
Lester Lu committed
601 602
  txfm_param.mode = mbmi->mode;
#endif  // CONFIG_LGT
Angie Chiang's avatar
Angie Chiang committed
603

Thomas Daede's avatar
Thomas Daede committed
604
#if !CONFIG_PVQ
605
  txfm_param.bd = xd->bd;
606
  const int is_hbd = get_bitdepth_data_path_index(xd);
Angie Chiang's avatar
Angie Chiang committed
607 608 609 610

#if CONFIG_TXMG
  av1_highbd_fwd_txfm(src_diff, coeff, diff_stride, &txfm_param);
#else   // CONFIG_TXMG
611
  fwd_txfm_func[is_hbd](src_diff, coeff, diff_stride, &txfm_param);
Angie Chiang's avatar
Angie Chiang committed
612
#endif  // CONFIG_TXMG
613

Yaowu Xu's avatar
Yaowu Xu committed
614
  if (xform_quant_idx != AV1_XFORM_QUANT_SKIP_QUANT) {
615
    if (LIKELY(!x->skip_block)) {
616
      quant_func_list[xform_quant_idx][is_hbd](
617
          coeff, tx2d_size, p, qcoeff, pd, dqcoeff, eob, scan_order, &qparam);
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
626 627
  return;
#else  // CONFIG_PVQ
Angie Chiang's avatar
Angie Chiang committed
628
  (void)xform_quant_idx;
629
#if CONFIG_HIGHBITDEPTH
630
  txfm_param.bd = xd->bd;
Thomas Daede's avatar
Thomas Daede committed
631
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
632 633
    av1_highbd_fwd_txfm(src_int16, coeff, diff_stride, &txfm_param);
    av1_highbd_fwd_txfm(pred, ref_coeff, diff_stride, &txfm_param);
Thomas Daede's avatar
Thomas Daede committed
634 635
  } else {
#endif
636 637
    av1_fwd_txfm(src_int16, coeff, diff_stride, &txfm_param);
    av1_fwd_txfm(pred, ref_coeff, diff_stride, &txfm_param);
638
#if CONFIG_HIGHBITDEPTH
Thomas Daede's avatar
Thomas Daede committed
639 640
  }
#endif
641 642

  // PVQ for inter mode block
643
  if (!x->skip_block) {
ltrudeau's avatar
ltrudeau committed
644
    PVQ_SKIP_TYPE ac_dc_coded =
Thomas Daede's avatar
Thomas Daede committed
645
        av1_pvq_encode_helper(x,
ltrudeau's avatar
ltrudeau committed
646 647 648 649 650 651 652 653 654 655 656 657
                              coeff,        // target original vector
                              ref_coeff,    // reference vector
                              dqcoeff,      // de-quantized vector
                              eob,          // End of Block marker
                              pd->dequant,  // aom's quantizers
                              plane,        // image plane
                              tx_size,      // block size in log_2 - 2
                              tx_type,
                              &x->rate,  // rate measured
                              x->pvq_speed,
                              pvq_info);  // PVQ info for a block
    skip = ac_dc_coded == PVQ_SKIP;
658
  }
659 660 661 662
  x->pvq_skip[plane] = skip;

  if (!skip) mbmi->skip = 0;
#endif  // #if !CONFIG_PVQ
Jingning Han's avatar
Jingning Han committed
663 664
}

665
static void encode_block(int plane, int block, int blk_row, int blk_col,
666
                         BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
Jingning Han's avatar
Jingning Han committed
667
  struct encode_b_args *const args = arg;
Angie Chiang's avatar
Angie Chiang committed
668
  AV1_COMMON *cm = args->cm;
Jingning Han's avatar
Jingning Han committed
669 670
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
671
  int ctx;
Jingning Han's avatar
Jingning Han committed
672 673 674
  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
675 676 677
#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
678
  uint8_t *dst;
679
#if !CONFIG_PVQ
Jingning Han's avatar
Jingning Han committed
680
  ENTROPY_CONTEXT *a, *l;
681
#endif
682
#if CONFIG_VAR_TX
683
  int bw = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
684
#endif
685 686
  dst = &pd->dst
             .buf[(blk_row * pd->dst.stride + blk_col) << tx_size_wide_log2[0]];
687 688

#if !CONFIG_PVQ
689 690
  a = &args->ta[blk_col];
  l = &args->tl[blk_row];
691 692 693 694 695
#if CONFIG_VAR_TX
  ctx = get_entropy_context(tx_size, a, l);
#else
  ctx = combine_entropy_contexts(*a, *l);
#endif
696 697 698
#else
  ctx = 0;
#endif  // CONFIG_PVQ
Jingning Han's avatar
Jingning Han committed
699

Alex Converse's avatar
Alex Converse committed
700
#if CONFIG_VAR_TX
701
  // Assert not magic number (uninitialized).
702
  assert(x->blk_skip[plane][blk_row * bw + blk_col] != 234);
703

704
  if (x->blk_skip[plane][blk_row * bw + blk_col] == 0)
705
#endif
706
  {
Angie Chiang's avatar
Angie Chiang committed
707
    av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
708
                    ctx, AV1_XFORM_QUANT_FP);
Jingning Han's avatar
Jingning Han committed
709
  }
710 711
#if CONFIG_VAR_TX
  else {
712
    p->eobs[block] = 0;
713 714
  }
#endif
715

716
#if !CONFIG_PVQ
Jingning Han's avatar
Jingning Han committed
717
  av1_optimize_b(cm, x, plane, blk_row, blk_col, block, plane_bsize, tx_size, a,
718
                 l, 0);
719

720 721
  av1_set_txb_context(x, plane, block, tx_size, a, l);

722
  if (p->eobs[block]) *(args->skip) = 0;
Jingning Han's avatar
Jingning Han committed
723

724
  if (p->eobs[block] != 0)
725 726 727 728
#else
  (void)ctx;
  if (!x->pvq_skip[plane]) *(args->skip) = 0;

729
  if (!x->pvq_skip[plane])
730
#endif
731
  {
Sarah Parker's avatar
Sarah Parker committed
732
#if CONFIG_LGT
733
    PREDICTION_MODE mode = xd->mi[0]->mbmi.mode;
Sarah Parker's avatar
Sarah Parker committed
734
#endif  // CONFIG_LGT
735 736 737
    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,
Lester Lu's avatar
Lester Lu committed
738
#if CONFIG_LGT
739
                                mode,
Sarah Parker's avatar
Sarah Parker committed
740 741
#endif  // CONFIG_LGT
#if CONFIG_MRC_TX && SIGNAL_ANY_MRC_MASK
742
                                mrc_mask,
Sarah Parker's avatar
Sarah Parker committed
743
#endif  // CONFIG_MRC_TX && SIGNAL_ANY_MRC_MASK
744 745 746
                                tx_type, tx_size, dst, pd->dst.stride,
                                p->eobs[block]);
  }
Jingning Han's avatar
Jingning Han committed
747 748
}

749 750 751 752 753 754 755 756
#if CONFIG_VAR_TX
static void encode_block_inter(int plane, int block, int blk_row, int blk_col,
                               BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
                               void *arg) {
  struct encode_b_args *const args = arg;
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
757
  const BLOCK_SIZE bsize = txsize_to_bsize[tx_size];
758
  const struct macroblockd_plane *const pd = &xd->plane[plane];
759 760
  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
761
  TX_SIZE plane_tx_size;
762 763
  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
764

765
  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
766

767 768 769
  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
770

771
  if (tx_size == plane_tx_size) {
772
    encode_block(plane, block, blk_row, blk_col, plane_bsize, tx_size, arg);
773
  } else {
774
    assert(tx_size < TX_SIZES_ALL);
Yue Chen's avatar
Yue Chen committed
775 776 777 778 779
#if CONFIG_RECT_TX_EXT
    int is_qttx = plane_tx_size == quarter_txsize_lookup[plane_bsize];
    const TX_SIZE sub_txs = is_qttx ? plane_tx_size : sub_tx_size_map[tx_size];
    if (is_qttx) assert(blk_row == 0 && blk_col == 0 && block == 0);
#else
780
    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
781 782
    assert(IMPLIES(tx_size <= TX_4X4, sub_txs == tx_size));
    assert(IMPLIES(tx_size > TX_4X4, sub_txs < tx_size));
Yue Chen's avatar
Yue Chen committed
783
#endif
784 785
    // This is the square transform block partition entry point.
    int bsl = tx_size_wide_unit[sub_txs];
786 787
    int i;
    assert(bsl > 0);
788

789
    for (i = 0; i < 4; ++i) {
Yue Chen's avatar
Yue Chen committed
790 791 792 793 794 795 796 797 798
#if CONFIG_RECT_TX_EXT
      int is_wide_tx = tx_size_wide_unit[sub_txs] > tx_size_high_unit[sub_txs];
      const int offsetr =
          is_qttx ? (is_wide_tx ? i * tx_size_high_unit[sub_txs] : 0)
                  : blk_row + ((i >> 1) * bsl);
      const int offsetc =
          is_qttx ? (is_wide_tx ? 0 : i * tx_size_wide_unit[sub_txs])
                  : blk_col + ((i & 0x01) * bsl);
#else
799 800
      const int offsetr = blk_row + ((i >> 1) * bsl);
      const int offsetc = blk_col + ((i & 0x01