vp9_encodemb.c 24.4 KB
Newer Older
John Koleszar's avatar
John Koleszar committed
1
/*
2
 *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
John Koleszar's avatar
John Koleszar committed
3
 *
4
 *  Use of this source code is governed by a BSD-style license
5
6
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
7
 *  in the file PATENTS.  All contributing project authors may
8
 *  be found in the AUTHORS file in the root of the source tree.
John Koleszar's avatar
John Koleszar committed
9
10
 */

11
12

#include "./vp9_rtcd.h"
13
#include "./vpx_config.h"
14
15
16
17

#include "vpx_mem/vpx_mem.h"

#include "vp9/common/vp9_idct.h"
18
19
20
#include "vp9/common/vp9_reconinter.h"
#include "vp9/common/vp9_reconintra.h"
#include "vp9/common/vp9_systemdependent.h"
21
22
23
24
25

#include "vp9/encoder/vp9_encodemb.h"
#include "vp9/encoder/vp9_quantize.h"
#include "vp9/encoder/vp9_rdopt.h"
#include "vp9/encoder/vp9_tokenize.h"
John Koleszar's avatar
John Koleszar committed
26

27
28
29
30
void vp9_subtract_block_c(int rows, int cols,
                          int16_t *diff_ptr, ptrdiff_t diff_stride,
                          const uint8_t *src_ptr, ptrdiff_t src_stride,
                          const uint8_t *pred_ptr, ptrdiff_t pred_stride) {
John Koleszar's avatar
John Koleszar committed
31
  int r, c;
John Koleszar's avatar
John Koleszar committed
32

33
34
  for (r = 0; r < rows; r++) {
    for (c = 0; c < cols; c++)
John Koleszar's avatar
John Koleszar committed
35
36
      diff_ptr[c] = src_ptr[c] - pred_ptr[c];

37
38
    diff_ptr += diff_stride;
    pred_ptr += pred_stride;
John Koleszar's avatar
John Koleszar committed
39
40
    src_ptr  += src_stride;
  }
John Koleszar's avatar
John Koleszar committed
41
42
}

43
static void subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
44
  struct macroblock_plane *const p = &x->plane[plane];
45
46
47
48
  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 = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
  const int bh = 4 * num_4x4_blocks_high_lookup[plane_bsize];
49

50
  vp9_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
51
                     pd->dst.buf, pd->dst.stride);
Yaowu Xu's avatar
Yaowu Xu committed
52
53
}

54
void vp9_subtract_sby(MACROBLOCK *x, BLOCK_SIZE bsize) {
55
  subtract_plane(x, bsize, 0);
56
57
}

58
void vp9_subtract_sbuv(MACROBLOCK *x, BLOCK_SIZE bsize) {
59
  int i;
60

61
62
  for (i = 1; i < MAX_MB_PLANE; i++)
    subtract_plane(x, bsize, i);
63
64
}

65
void vp9_subtract_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
66
67
  vp9_subtract_sby(x, bsize);
  vp9_subtract_sbuv(x, bsize);
John Koleszar's avatar
John Koleszar committed
68
69
}

70
#define RDTRUNC(RM, DM, R, D) ((128 + (R) * (RM)) & 0xFF)
71
typedef struct vp9_token_state vp9_token_state;
72

73
struct vp9_token_state {
74
75
  int           rate;
  int           error;
Daniel Kang's avatar
Daniel Kang committed
76
  int           next;
77
78
79
80
  signed char   token;
  short         qc;
};

81
// TODO(jimbankoski): experiment to find optimal RD numbers.
82
83
#define Y1_RD_MULT 4
#define UV_RD_MULT 2
84

John Koleszar's avatar
John Koleszar committed
85
86
87
static const int plane_rd_mult[4] = {
  Y1_RD_MULT,
  UV_RD_MULT,
88
89
};

90
91
92
93
94
95
96
97
98
99
#define UPDATE_RD_COST()\
{\
  rd_cost0 = RDCOST(rdmult, rddiv, rate0, error0);\
  rd_cost1 = RDCOST(rdmult, rddiv, rate1, error1);\
  if (rd_cost0 == rd_cost1) {\
    rd_cost0 = RDTRUNC(rdmult, rddiv, rate0, error0);\
    rd_cost1 = RDTRUNC(rdmult, rddiv, rate1, error1);\
  }\
}

100
101
// This function is a place holder for now but may ultimately need
// to scan previous tokens to work out the correct context.
102
103
static int trellis_get_coeff_context(const int16_t *scan,
                                     const int16_t *nb,
104
                                     int idx, int token,
105
                                     uint8_t *token_cache) {
106
  int bak = token_cache[scan[idx]], pt;
107
  token_cache[scan[idx]] = vp9_pt_energy_class[token];
108
  pt = get_coef_context(nb, token_cache, idx + 1);
109
  token_cache[scan[idx]] = bak;
110
  return pt;
111
112
}

113
static void optimize_b(MACROBLOCK *mb,
114
                       int plane, int block, BLOCK_SIZE plane_bsize,
115
                       ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
John Koleszar's avatar
John Koleszar committed
116
                       TX_SIZE tx_size) {
117
  MACROBLOCKD *const xd = &mb->e_mbd;
118
  struct macroblock_plane *p = &mb->plane[plane];
119
  struct macroblockd_plane *pd = &xd->plane[plane];
120
  const int ref = is_inter_block(&xd->mi_8x8[0]->mbmi);
121
122
  vp9_token_state tokens[1025][2];
  unsigned best_index[1025][2];
123
  const int16_t *coeff_ptr = BLOCK_OFFSET(mb->plane[plane].coeff, block);
124
125
  int16_t *qcoeff_ptr;
  int16_t *dqcoeff_ptr;
126
  int eob = p->eobs[block], final_eob, sz = 0;
127
  const int i0 = 0;
128
  int rc, x, next, i;
129
130
131
  int64_t rdmult, rddiv, rd_cost0, rd_cost1;
  int rate0, rate1, error0, error1, t0, t1;
  int best, band, pt;
132
  PLANE_TYPE type = pd->plane_type;
John Koleszar's avatar
John Koleszar committed
133
  int err_mult = plane_rd_mult[type];
134
  const int default_eob = 16 << (tx_size << 1);
135

136
  const int mul = 1 + (tx_size == TX_32X32);
137
  uint8_t token_cache[1024];
138
  const int16_t *dequant_ptr = pd->dequant;
139
  const uint8_t *const band_translate = get_band_translate(tx_size);
140
141
142
  const scan_order *so = get_scan(xd, tx_size, type, block);
  const int16_t *scan = so->scan;
  const int16_t *nb = so->neighbors;
John Koleszar's avatar
John Koleszar committed
143

John Koleszar's avatar
John Koleszar committed
144
  assert((!type && !plane) || (type && plane));
145
  dqcoeff_ptr = BLOCK_OFFSET(pd->dqcoeff, block);
146
  qcoeff_ptr = BLOCK_OFFSET(p->qcoeff, block);
John Koleszar's avatar
John Koleszar committed
147
  assert(eob <= default_eob);
John Koleszar's avatar
John Koleszar committed
148
149
150

  /* Now set up a Viterbi trellis to evaluate alternative roundings. */
  rdmult = mb->rdmult * err_mult;
151
  if (!is_inter_block(&mb->e_mbd.mi_8x8[0]->mbmi))
John Koleszar's avatar
John Koleszar committed
152
153
154
155
156
    rdmult = (rdmult * 9) >> 4;
  rddiv = mb->rddiv;
  /* Initialize the sentinel node of the trellis. */
  tokens[eob][0].rate = 0;
  tokens[eob][0].error = 0;
157
  tokens[eob][0].next = default_eob;
158
  tokens[eob][0].token = EOB_TOKEN;
John Koleszar's avatar
John Koleszar committed
159
160
161
  tokens[eob][0].qc = 0;
  *(tokens[eob] + 1) = *(tokens[eob] + 0);
  next = eob;
162
  for (i = 0; i < eob; i++)
163
164
    token_cache[scan[i]] = vp9_pt_energy_class[vp9_dct_value_tokens_ptr[
        qcoeff_ptr[scan[i]]].token];
165

John Koleszar's avatar
John Koleszar committed
166
  for (i = eob; i-- > i0;) {
167
    int base_bits, d2, dx;
John Koleszar's avatar
John Koleszar committed
168

169
    rc = scan[i];
John Koleszar's avatar
John Koleszar committed
170
171
172
173
174
175
176
177
178
    x = qcoeff_ptr[rc];
    /* Only add a trellis state for non-zero coefficients. */
    if (x) {
      int shortcut = 0;
      error0 = tokens[next][0].error;
      error1 = tokens[next][1].error;
      /* Evaluate the first possibility for this state. */
      rate0 = tokens[next][0].rate;
      rate1 = tokens[next][1].rate;
179
      t0 = (vp9_dct_value_tokens_ptr + x)->token;
John Koleszar's avatar
John Koleszar committed
180
      /* Consider both possible successor states. */
181
      if (next < default_eob) {
182
        band = band_translate[i + 1];
183
        pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
John Koleszar's avatar
John Koleszar committed
184
        rate0 +=
185
          mb->token_costs[tx_size][type][ref][band][0][pt]
186
                         [tokens[next][0].token];
John Koleszar's avatar
John Koleszar committed
187
        rate1 +=
188
          mb->token_costs[tx_size][type][ref][band][0][pt]
189
                         [tokens[next][1].token];
John Koleszar's avatar
John Koleszar committed
190
      }
191
      UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
192
193
      /* And pick the best. */
      best = rd_cost1 < rd_cost0;
194
      base_bits = *(vp9_dct_value_cost_ptr + x);
195
      dx = mul * (dqcoeff_ptr[rc] - coeff_ptr[rc]);
John Koleszar's avatar
John Koleszar committed
196
197
198
199
200
201
      d2 = dx * dx;
      tokens[i][0].rate = base_bits + (best ? rate1 : rate0);
      tokens[i][0].error = d2 + (best ? error1 : error0);
      tokens[i][0].next = next;
      tokens[i][0].token = t0;
      tokens[i][0].qc = x;
202
      best_index[i][0] = best;
203

John Koleszar's avatar
John Koleszar committed
204
205
206
207
      /* Evaluate the second possibility for this state. */
      rate0 = tokens[next][0].rate;
      rate1 = tokens[next][1].rate;

208
209
210
      if ((abs(x)*dequant_ptr[rc != 0] > abs(coeff_ptr[rc]) * mul) &&
          (abs(x)*dequant_ptr[rc != 0] < abs(coeff_ptr[rc]) * mul +
                                         dequant_ptr[rc != 0]))
John Koleszar's avatar
John Koleszar committed
211
212
213
214
215
216
217
218
219
220
221
222
223
        shortcut = 1;
      else
        shortcut = 0;

      if (shortcut) {
        sz = -(x < 0);
        x -= 2 * sz + 1;
      }

      /* Consider both possible successor states. */
      if (!x) {
        /* If we reduced this coefficient to zero, check to see if
         *  we need to move the EOB back here.
224
         */
225
226
        t0 = tokens[next][0].token == EOB_TOKEN ? EOB_TOKEN : ZERO_TOKEN;
        t1 = tokens[next][1].token == EOB_TOKEN ? EOB_TOKEN : ZERO_TOKEN;
John Koleszar's avatar
John Koleszar committed
227
      } else {
228
        t0 = t1 = (vp9_dct_value_tokens_ptr + x)->token;
John Koleszar's avatar
John Koleszar committed
229
      }
230
      if (next < default_eob) {
231
        band = band_translate[i + 1];
232
        if (t0 != EOB_TOKEN) {
233
          pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
234
          rate0 += mb->token_costs[tx_size][type][ref][band][!x][pt]
235
                                  [tokens[next][0].token];
John Koleszar's avatar
John Koleszar committed
236
        }
237
        if (t1 != EOB_TOKEN) {
238
          pt = trellis_get_coeff_context(scan, nb, i, t1, token_cache);
239
          rate1 += mb->token_costs[tx_size][type][ref][band][!x][pt]
240
                                  [tokens[next][1].token];
John Koleszar's avatar
John Koleszar committed
241
242
        }
      }
John Koleszar's avatar
John Koleszar committed
243

244
      UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
245
246
      /* And pick the best. */
      best = rd_cost1 < rd_cost0;
247
      base_bits = *(vp9_dct_value_cost_ptr + x);
John Koleszar's avatar
John Koleszar committed
248
249

      if (shortcut) {
250
        dx -= (dequant_ptr[rc != 0] + sz) ^ sz;
John Koleszar's avatar
John Koleszar committed
251
252
253
254
255
256
257
        d2 = dx * dx;
      }
      tokens[i][1].rate = base_bits + (best ? rate1 : rate0);
      tokens[i][1].error = d2 + (best ? error1 : error0);
      tokens[i][1].next = next;
      tokens[i][1].token = best ? t1 : t0;
      tokens[i][1].qc = x;
258
      best_index[i][1] = best;
John Koleszar's avatar
John Koleszar committed
259
260
      /* Finally, make this the new head of the trellis. */
      next = i;
261
262
263
264
    } else {
      /* There's no choice to make for a zero coefficient, so we don't
       *  add a new trellis node, but we do need to update the costs.
       */
265
      band = band_translate[i + 1];
John Koleszar's avatar
John Koleszar committed
266
267
268
      t0 = tokens[next][0].token;
      t1 = tokens[next][1].token;
      /* Update the cost of each path if we're past the EOB token. */
269
      if (t0 != EOB_TOKEN) {
270
        tokens[next][0].rate +=
271
            mb->token_costs[tx_size][type][ref][band][1][0][t0];
John Koleszar's avatar
John Koleszar committed
272
273
        tokens[next][0].token = ZERO_TOKEN;
      }
274
      if (t1 != EOB_TOKEN) {
275
        tokens[next][1].rate +=
276
            mb->token_costs[tx_size][type][ref][band][1][0][t1];
John Koleszar's avatar
John Koleszar committed
277
278
        tokens[next][1].token = ZERO_TOKEN;
      }
279
      best_index[i][0] = best_index[i][1] = 0;
John Koleszar's avatar
John Koleszar committed
280
      /* Don't update next, because we didn't add a new node. */
281
    }
John Koleszar's avatar
John Koleszar committed
282
283
284
  }

  /* Now pick the best path through the whole trellis. */
285
  band = band_translate[i + 1];
286
  pt = combine_entropy_contexts(*a, *l);
John Koleszar's avatar
John Koleszar committed
287
288
289
290
291
292
  rate0 = tokens[next][0].rate;
  rate1 = tokens[next][1].rate;
  error0 = tokens[next][0].error;
  error1 = tokens[next][1].error;
  t0 = tokens[next][0].token;
  t1 = tokens[next][1].token;
293
294
  rate0 += mb->token_costs[tx_size][type][ref][band][0][pt][t0];
  rate1 += mb->token_costs[tx_size][type][ref][band][0][pt][t1];
295
  UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
296
297
  best = rd_cost1 < rd_cost0;
  final_eob = i0 - 1;
298
299
  vpx_memset(qcoeff_ptr, 0, sizeof(*qcoeff_ptr) * (16 << (tx_size * 2)));
  vpx_memset(dqcoeff_ptr, 0, sizeof(*dqcoeff_ptr) * (16 << (tx_size * 2)));
John Koleszar's avatar
John Koleszar committed
300
301
  for (i = next; i < eob; i = next) {
    x = tokens[i][best].qc;
302
    if (x) {
John Koleszar's avatar
John Koleszar committed
303
      final_eob = i;
304
    }
305
    rc = scan[i];
John Koleszar's avatar
John Koleszar committed
306
    qcoeff_ptr[rc] = x;
307
    dqcoeff_ptr[rc] = (x * dequant_ptr[rc != 0]) / mul;
308

John Koleszar's avatar
John Koleszar committed
309
    next = tokens[i][best].next;
310
    best = best_index[i][best];
John Koleszar's avatar
John Koleszar committed
311
312
313
  }
  final_eob++;

314
  mb->plane[plane].eobs[block] = final_eob;
315
  *a = *l = (final_eob > 0);
316
317
}

318
void vp9_optimize_b(int plane, int block, BLOCK_SIZE plane_bsize,
319
                    TX_SIZE tx_size, MACROBLOCK *mb, struct optimize_ctx *ctx) {
John Koleszar's avatar
John Koleszar committed
320
  int x, y;
321
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x, &y);
322
323
  optimize_b(mb, plane, block, plane_bsize,
             &ctx->ta[plane][x], &ctx->tl[plane][y], tx_size);
324
325
}

326
static void optimize_init_b(int plane, BLOCK_SIZE bsize,
327
                            struct encode_b_args *args) {
328
329
  const MACROBLOCKD *xd = &args->x->e_mbd;
  const struct macroblockd_plane* const pd = &xd->plane[plane];
330
  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
331
332
  const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
  const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
333
  const MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
334
  const TX_SIZE tx_size = plane ? get_uv_tx_size(mbmi) : mbmi->tx_size;
335

336
337
338
  vp9_get_entropy_contexts(tx_size, args->ctx->ta[plane], args->ctx->tl[plane],
                           pd->above_context, pd->left_context,
                           num_4x4_w, num_4x4_h);
339
}
340
void vp9_xform_quant(int plane, int block, BLOCK_SIZE plane_bsize,
341
                     TX_SIZE tx_size, void *arg) {
342
343
344
  struct encode_b_args* const args = arg;
  MACROBLOCK* const x = args->x;
  MACROBLOCKD* const xd = &x->e_mbd;
345
346
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
347
  int16_t *coeff = BLOCK_OFFSET(p->coeff, block);
348
  int16_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
349
  int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
350
  const scan_order *scan_order;
351
  uint16_t *eob = &p->eobs[block];
352
353
  const int diff_stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
  int i, j;
354
  int16_t *src_diff;
355
356
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
  src_diff = &p->src_diff[4 * (j * diff_stride + i)];
357
358

  switch (tx_size) {
359
    case TX_32X32:
360
      scan_order = &vp9_default_scan_orders[TX_32X32];
361
      if (x->use_lp32x32fdct)
362
        vp9_fdct32x32_rd(src_diff, coeff, diff_stride);
363
      else
364
        vp9_fdct32x32(src_diff, coeff, diff_stride);
365
366
      vp9_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round,
                           p->quant, p->quant_shift, qcoeff, dqcoeff,
367
368
                           pd->dequant, p->zbin_extra, eob, scan_order->scan,
                           scan_order->iscan);
369
370
      break;
    case TX_16X16:
371
      scan_order = &vp9_default_scan_orders[TX_16X16];
372
      vp9_fdct16x16(src_diff, coeff, diff_stride);
373
374
      vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
                     p->quant, p->quant_shift, qcoeff, dqcoeff,
375
376
                     pd->dequant, p->zbin_extra, eob,
                     scan_order->scan, scan_order->iscan);
377
378
      break;
    case TX_8X8:
379
      scan_order = &vp9_default_scan_orders[TX_8X8];
380
      vp9_fdct8x8(src_diff, coeff, diff_stride);
381
382
      vp9_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round,
                     p->quant, p->quant_shift, qcoeff, dqcoeff,
383
384
                     pd->dequant, p->zbin_extra, eob,
                     scan_order->scan, scan_order->iscan);
385
386
      break;
    case TX_4X4:
387
      scan_order = &vp9_default_scan_orders[TX_4X4];
388
      x->fwd_txm4x4(src_diff, coeff, diff_stride);
389
390
      vp9_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round,
                     p->quant, p->quant_shift, qcoeff, dqcoeff,
391
392
                     pd->dequant, p->zbin_extra, eob,
                     scan_order->scan, scan_order->iscan);
393
394
395
      break;
    default:
      assert(0);
John Koleszar's avatar
John Koleszar committed
396
  }
397
398
}

399
static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize,
400
                         TX_SIZE tx_size, void *arg) {
401
402
403
  struct encode_b_args *const args = arg;
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
404
  struct optimize_ctx *const ctx = args->ctx;
405
  struct macroblock_plane *const p = &x->plane[plane];
406
  struct macroblockd_plane *const pd = &xd->plane[plane];
407
  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
408
409
410
411
  int i, j;
  uint8_t *dst;
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
  dst = &pd->dst.buf[4 * j * pd->dst.stride + 4 * i];
412
413
414
415

  // TODO(jingning): per transformed block zero forcing only enabled for
  // luma component. will integrate chroma components as well.
  if (x->zcoeff_blk[tx_size][block] && plane == 0) {
416
    p->eobs[block] = 0;
417
    ctx->ta[plane][i] = 0;
418
    ctx->tl[plane][j] = 0;
419
420
421
    return;
  }

422
  if (!x->skip_recode)
423
    vp9_xform_quant(plane, block, plane_bsize, tx_size, arg);
424

425
  if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
426
    vp9_optimize_b(plane, block, plane_bsize, tx_size, x, ctx);
427
  } else {
428
429
    ctx->ta[plane][i] = p->eobs[block] > 0;
    ctx->tl[plane][j] = p->eobs[block] > 0;
430
  }
John Koleszar's avatar
John Koleszar committed
431

Jim Bankoski's avatar
Jim Bankoski committed
432
433
434
  if (p->eobs[block])
    *(args->skip_coeff) = 0;

435
  if (x->skip_encode || p->eobs[block] == 0)
436
    return;
437

438
  switch (tx_size) {
439
    case TX_32X32:
440
      vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
441
442
      break;
    case TX_16X16:
443
      vp9_idct16x16_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
444
445
      break;
    case TX_8X8:
446
      vp9_idct8x8_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
447
448
      break;
    case TX_4X4:
449
450
451
      // this is like vp9_short_idct4x4 but has a special case around eob<=1
      // which is significant (not just an optimization) for the lossless
      // case.
452
      xd->itxm_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
453
      break;
454
    default:
James Zern's avatar
James Zern committed
455
      assert(0 && "Invalid transform size");
456
  }
John Koleszar's avatar
John Koleszar committed
457
}
458
459
460
461
462
static void encode_block_pass1(int plane, int block, 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;
463
  struct macroblock_plane *const p = &x->plane[plane];
464
465
  struct macroblockd_plane *const pd = &xd->plane[plane];
  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
466
467
468
469
  int i, j;
  uint8_t *dst;
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
  dst = &pd->dst.buf[4 * j * pd->dst.stride + 4 * i];
470
471
472

  vp9_xform_quant(plane, block, plane_bsize, tx_size, arg);

473
  if (p->eobs[block] == 0)
474
475
    return;

476
  xd->itxm_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
477
478
}

479
void vp9_encode_sby(MACROBLOCK *x, BLOCK_SIZE bsize) {
480
  MACROBLOCKD *const xd = &x->e_mbd;
481
  struct optimize_ctx ctx;
Jim Bankoski's avatar
Jim Bankoski committed
482
483
  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
  struct encode_b_args arg = {x, &ctx, &mbmi->skip_coeff};
484
485
486

  vp9_subtract_sby(x, bsize);
  if (x->optimize)
487
    optimize_init_b(0, bsize, &arg);
488

489
490
  vp9_foreach_transformed_block_in_plane(xd, bsize, 0, encode_block_pass1,
                                         &arg);
491
492
}

493
void vp9_encode_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
494
  MACROBLOCKD *const xd = &x->e_mbd;
495
  struct optimize_ctx ctx;
Jim Bankoski's avatar
Jim Bankoski committed
496
497
  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
  struct encode_b_args arg = {x, &ctx, &mbmi->skip_coeff};
John Koleszar's avatar
John Koleszar committed
498

499
  if (!x->skip_recode)
500
    vp9_subtract_sb(x, bsize);
501

502
  if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
503
504
505
506
    int i;
    for (i = 0; i < MAX_MB_PLANE; ++i)
      optimize_init_b(i, bsize, &arg);
  }
Scott LaVarnway's avatar
Scott LaVarnway committed
507

508
  vp9_foreach_transformed_block(xd, bsize, encode_block, &arg);
John Koleszar's avatar
John Koleszar committed
509
}
510

511
void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
512
                            TX_SIZE tx_size, void *arg) {
513
  struct encode_b_args* const args = arg;
514
515
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
516
  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
517
518
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
519
  int16_t *coeff = BLOCK_OFFSET(p->coeff, block);
520
  int16_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
521
  int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
522
  const scan_order *scan_order;
523
  TX_TYPE tx_type;
524
  MB_PREDICTION_MODE mode;
525
526
  const int bwl = b_width_log2(plane_bsize);
  const int diff_stride = 4 * (1 << bwl);
527
528
  uint8_t *src, *dst;
  int16_t *src_diff;
529
  uint16_t *eob = &p->eobs[block];
530
531
532
533
534
  int i, j;
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
  dst = &pd->dst.buf[4 * (j * pd->dst.stride + i)];
  src = &p->src.buf[4 * (j * p->src.stride + i)];
  src_diff = &p->src_diff[4 * (j * diff_stride + i)];
535

536
  // if (x->optimize)
537
  // vp9_optimize_b(plane, block, plane_bsize, tx_size, x, args->ctx);
538

539
  switch (tx_size) {
540
    case TX_32X32:
541
      scan_order = &vp9_default_scan_orders[TX_32X32];
542
      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
543
      vp9_predict_intra_block(xd, block >> 6, bwl, TX_32X32, mode,
544
545
                              x->skip_encode ? src : dst,
                              x->skip_encode ? p->src.stride : pd->dst.stride,
546
                              dst, pd->dst.stride, i, j, plane);
547
      if (!x->skip_recode) {
548
        vp9_subtract_block(32, 32, src_diff, diff_stride,
549
550
                           src, p->src.stride, dst, pd->dst.stride);
        if (x->use_lp32x32fdct)
551
          vp9_fdct32x32_rd(src_diff, coeff, diff_stride);
552
        else
553
          vp9_fdct32x32(src_diff, coeff, diff_stride);
554
555
        vp9_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round,
                             p->quant, p->quant_shift, qcoeff, dqcoeff,
556
557
                             pd->dequant, p->zbin_extra, eob, scan_order->scan,
                             scan_order->iscan);
558
      }
559
      if (!x->skip_encode && *eob)
Yaowu Xu's avatar
Yaowu Xu committed
560
        vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, *eob);
561
562
      break;
    case TX_16X16:
563
      tx_type = get_tx_type_16x16(pd->plane_type, xd);
564
      scan_order = &vp9_scan_orders[TX_16X16][tx_type];
565
      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
566
      vp9_predict_intra_block(xd, block >> 4, bwl, TX_16X16, mode,
567
568
                              x->skip_encode ? src : dst,
                              x->skip_encode ? p->src.stride : pd->dst.stride,
569
                              dst, pd->dst.stride, i, j, plane);
570
      if (!x->skip_recode) {
571
        vp9_subtract_block(16, 16, src_diff, diff_stride,
572
                           src, p->src.stride, dst, pd->dst.stride);
573
        vp9_fht16x16(src_diff, coeff, diff_stride, tx_type);
574
575
        vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
                       p->quant, p->quant_shift, qcoeff, dqcoeff,
576
577
                       pd->dequant, p->zbin_extra, eob, scan_order->scan,
                       scan_order->iscan);
578
      }
579
      if (!x->skip_encode && *eob)
580
        vp9_iht16x16_add(tx_type, dqcoeff, dst, pd->dst.stride, *eob);
581
582
      break;
    case TX_8X8:
583
      tx_type = get_tx_type_8x8(pd->plane_type, xd);
584
      scan_order = &vp9_scan_orders[TX_8X8][tx_type];
585
      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
586
      vp9_predict_intra_block(xd, block >> 2, bwl, TX_8X8, mode,
587
588
                              x->skip_encode ? src : dst,
                              x->skip_encode ? p->src.stride : pd->dst.stride,
589
                              dst, pd->dst.stride, i, j, plane);
590
      if (!x->skip_recode) {
591
        vp9_subtract_block(8, 8, src_diff, diff_stride,
592
                           src, p->src.stride, dst, pd->dst.stride);
593
        vp9_fht8x8(src_diff, coeff, diff_stride, tx_type);
594
595
        vp9_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round, p->quant,
                       p->quant_shift, qcoeff, dqcoeff,
596
597
                       pd->dequant, p->zbin_extra, eob, scan_order->scan,
                       scan_order->iscan);
598
      }
599
      if (!x->skip_encode && *eob)
600
        vp9_iht8x8_add(tx_type, dqcoeff, dst, pd->dst.stride, *eob);
601
602
      break;
    case TX_4X4:
603
      tx_type = get_tx_type_4x4(pd->plane_type, xd, block);
604
      scan_order = &vp9_scan_orders[TX_4X4][tx_type];
605
      if (mbmi->sb_type < BLOCK_8X8 && plane == 0)
606
        mode = xd->mi_8x8[0]->bmi[block].as_mode;
607
      else
608
        mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
609

610
      vp9_predict_intra_block(xd, block, bwl, TX_4X4, mode,
611
612
                              x->skip_encode ? src : dst,
                              x->skip_encode ? p->src.stride : pd->dst.stride,
613
                              dst, pd->dst.stride, i, j, plane);
614
615

      if (!x->skip_recode) {
616
        vp9_subtract_block(4, 4, src_diff, diff_stride,
617
618
                           src, p->src.stride, dst, pd->dst.stride);
        if (tx_type != DCT_DCT)
619
          vp9_fht4x4(src_diff, coeff, diff_stride, tx_type);
620
        else
621
          x->fwd_txm4x4(src_diff, coeff, diff_stride);
622
623
        vp9_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round, p->quant,
                       p->quant_shift, qcoeff, dqcoeff,
624
625
                       pd->dequant, p->zbin_extra, eob, scan_order->scan,
                       scan_order->iscan);
626
627
      }

628
      if (!x->skip_encode && *eob) {
629
630
631
632
        if (tx_type == DCT_DCT)
          // this is like vp9_short_idct4x4 but has a special case around eob<=1
          // which is significant (not just an optimization) for the lossless
          // case.
633
          xd->itxm_add(dqcoeff, dst, pd->dst.stride, *eob);
634
        else
635
          vp9_iht4x4_16_add(dqcoeff, dst, pd->dst.stride, tx_type);
636
      }
637
      break;
638
639
    default:
      assert(0);
640
  }
Jim Bankoski's avatar
Jim Bankoski committed
641
642
  if (*eob)
    *(args->skip_coeff) = 0;
643
644
}

645
void vp9_encode_intra_block_y(MACROBLOCK *x, BLOCK_SIZE bsize) {
646
647
  MACROBLOCKD* const xd = &x->e_mbd;
  struct optimize_ctx ctx;
Jim Bankoski's avatar
Jim Bankoski committed
648
649
  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
  struct encode_b_args arg = {x, &ctx, &mbmi->skip_coeff};
650

651
  vp9_foreach_transformed_block_in_plane(xd, bsize, 0, vp9_encode_block_intra,
652
                                     &arg);
653
}
654
void vp9_encode_intra_block_uv(MACROBLOCK *x, BLOCK_SIZE bsize) {
655
656
  MACROBLOCKD* const xd = &x->e_mbd;
  struct optimize_ctx ctx;
Jim Bankoski's avatar
Jim Bankoski committed
657
658
  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
  struct encode_b_args arg = {x, &ctx, &mbmi->skip_coeff};
659
  vp9_foreach_transformed_block_uv(xd, bsize, vp9_encode_block_intra, &arg);
660
}
661

662
663
664
665
666
667
668
669
670
671
672
int vp9_encode_intra(MACROBLOCK *x, int use_16x16_pred) {
  MB_MODE_INFO * mbmi = &x->e_mbd.mi_8x8[0]->mbmi;
  x->skip_encode = 0;
  mbmi->mode = DC_PRED;
  mbmi->ref_frame[0] = INTRA_FRAME;
  mbmi->tx_size = use_16x16_pred ? (mbmi->sb_type >= BLOCK_16X16 ? TX_16X16
                                                                 : TX_8X8)
                                   : TX_4X4;
  vp9_encode_intra_block_y(x, mbmi->sb_type);
  return vp9_get_mb_ss(x->plane[0].src_diff);
}