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
#include "vp9/encoder/vp9_dct.h"
23
24
25
26
#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
27

28
29
30
31
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
32
  int r, c;
John Koleszar's avatar
John Koleszar committed
33

34
35
  for (r = 0; r < rows; r++) {
    for (c = 0; c < cols; c++)
John Koleszar's avatar
John Koleszar committed
36
37
      diff_ptr[c] = src_ptr[c] - pred_ptr[c];

38
39
    diff_ptr += diff_stride;
    pred_ptr += pred_stride;
John Koleszar's avatar
John Koleszar committed
40
41
    src_ptr  += src_stride;
  }
John Koleszar's avatar
John Koleszar committed
42
43
}

44
static void subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
45
  struct macroblock_plane *const p = &x->plane[plane];
46
47
48
49
  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];
50

51
  vp9_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
52
                     pd->dst.buf, pd->dst.stride);
Yaowu Xu's avatar
Yaowu Xu committed
53
54
}

55
void vp9_subtract_sby(MACROBLOCK *x, BLOCK_SIZE bsize) {
56
  subtract_plane(x, bsize, 0);
57
58
}

59
void vp9_subtract_sbuv(MACROBLOCK *x, BLOCK_SIZE bsize) {
60
  int i;
61

62
63
  for (i = 1; i < MAX_MB_PLANE; i++)
    subtract_plane(x, bsize, i);
64
65
}

66
void vp9_subtract_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
67
68
  vp9_subtract_sby(x, bsize);
  vp9_subtract_sbuv(x, bsize);
John Koleszar's avatar
John Koleszar committed
69
70
}

71
#define RDTRUNC(RM, DM, R, D) ((128 + (R) * (RM)) & 0xFF)
72
typedef struct vp9_token_state vp9_token_state;
73

74
struct vp9_token_state {
75
76
  int           rate;
  int           error;
Daniel Kang's avatar
Daniel Kang committed
77
  int           next;
78
79
80
81
  signed char   token;
  short         qc;
};

82
// TODO(jimbankoski): experiment to find optimal RD numbers.
83
84
#define Y1_RD_MULT 4
#define UV_RD_MULT 2
85

John Koleszar's avatar
John Koleszar committed
86
87
88
static const int plane_rd_mult[4] = {
  Y1_RD_MULT,
  UV_RD_MULT,
89
90
};

91
92
93
94
95
96
97
98
99
100
#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);\
  }\
}

101
102
// This function is a place holder for now but may ultimately need
// to scan previous tokens to work out the correct context.
103
104
static int trellis_get_coeff_context(const int16_t *scan,
                                     const int16_t *nb,
105
                                     int idx, int token,
106
                                     uint8_t *token_cache) {
107
  int bak = token_cache[scan[idx]], pt;
108
  token_cache[scan[idx]] = vp9_pt_energy_class[token];
109
  pt = get_coef_context(nb, token_cache, idx + 1);
110
  token_cache[scan[idx]] = bak;
111
  return pt;
112
113
}

114
static void optimize_b(MACROBLOCK *mb,
115
                       int plane, int block, BLOCK_SIZE plane_bsize,
116
                       ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
John Koleszar's avatar
John Koleszar committed
117
                       TX_SIZE tx_size) {
118
  MACROBLOCKD *const xd = &mb->e_mbd;
119
  struct macroblock_plane *p = &mb->plane[plane];
120
  struct macroblockd_plane *pd = &xd->plane[plane];
121
  const int ref = is_inter_block(&xd->mi_8x8[0]->mbmi);
122
123
  vp9_token_state tokens[1025][2];
  unsigned best_index[1025][2];
124
  const int16_t *coeff_ptr = BLOCK_OFFSET(mb->plane[plane].coeff, block);
125
126
  int16_t *qcoeff_ptr;
  int16_t *dqcoeff_ptr;
127
  int eob = p->eobs[block], final_eob, sz = 0;
128
  const int i0 = 0;
129
  int rc, x, next, i;
130
131
132
  int64_t rdmult, rddiv, rd_cost0, rd_cost1;
  int rate0, rate1, error0, error1, t0, t1;
  int best, band, pt;
133
  PLANE_TYPE type = pd->plane_type;
John Koleszar's avatar
John Koleszar committed
134
  int err_mult = plane_rd_mult[type];
135
  const int default_eob = 16 << (tx_size << 1);
136

137
  const int mul = 1 + (tx_size == TX_32X32);
138
  uint8_t token_cache[1024];
139
  const int16_t *dequant_ptr = pd->dequant;
140
  const uint8_t *const band_translate = get_band_translate(tx_size);
141
142
143
  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
144

John Koleszar's avatar
John Koleszar committed
145
  assert((!type && !plane) || (type && plane));
146
  dqcoeff_ptr = BLOCK_OFFSET(pd->dqcoeff, block);
147
  qcoeff_ptr = BLOCK_OFFSET(p->qcoeff, block);
John Koleszar's avatar
John Koleszar committed
148
  assert(eob <= default_eob);
John Koleszar's avatar
John Koleszar committed
149
150
151

  /* Now set up a Viterbi trellis to evaluate alternative roundings. */
  rdmult = mb->rdmult * err_mult;
152
  if (!is_inter_block(&mb->e_mbd.mi_8x8[0]->mbmi))
John Koleszar's avatar
John Koleszar committed
153
154
155
156
157
    rdmult = (rdmult * 9) >> 4;
  rddiv = mb->rddiv;
  /* Initialize the sentinel node of the trellis. */
  tokens[eob][0].rate = 0;
  tokens[eob][0].error = 0;
158
  tokens[eob][0].next = default_eob;
159
  tokens[eob][0].token = EOB_TOKEN;
John Koleszar's avatar
John Koleszar committed
160
161
162
  tokens[eob][0].qc = 0;
  *(tokens[eob] + 1) = *(tokens[eob] + 0);
  next = eob;
163
  for (i = 0; i < eob; i++)
164
165
    token_cache[scan[i]] = vp9_pt_energy_class[vp9_dct_value_tokens_ptr[
        qcoeff_ptr[scan[i]]].token];
166

John Koleszar's avatar
John Koleszar committed
167
  for (i = eob; i-- > i0;) {
168
    int base_bits, d2, dx;
John Koleszar's avatar
John Koleszar committed
169

170
    rc = scan[i];
John Koleszar's avatar
John Koleszar committed
171
172
173
174
175
176
177
178
179
    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;
180
      t0 = (vp9_dct_value_tokens_ptr + x)->token;
John Koleszar's avatar
John Koleszar committed
181
      /* Consider both possible successor states. */
182
      if (next < default_eob) {
183
        band = band_translate[i + 1];
184
        pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
John Koleszar's avatar
John Koleszar committed
185
        rate0 +=
186
          mb->token_costs[tx_size][type][ref][band][0][pt]
187
                         [tokens[next][0].token];
John Koleszar's avatar
John Koleszar committed
188
        rate1 +=
189
          mb->token_costs[tx_size][type][ref][band][0][pt]
190
                         [tokens[next][1].token];
John Koleszar's avatar
John Koleszar committed
191
      }
192
      UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
193
194
      /* And pick the best. */
      best = rd_cost1 < rd_cost0;
195
      base_bits = *(vp9_dct_value_cost_ptr + x);
196
      dx = mul * (dqcoeff_ptr[rc] - coeff_ptr[rc]);
John Koleszar's avatar
John Koleszar committed
197
198
199
200
201
202
      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;
203
      best_index[i][0] = best;
204

John Koleszar's avatar
John Koleszar committed
205
206
207
208
      /* Evaluate the second possibility for this state. */
      rate0 = tokens[next][0].rate;
      rate1 = tokens[next][1].rate;

209
210
211
      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
212
213
214
215
216
217
218
219
220
221
222
223
224
        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.
225
         */
226
227
        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
228
      } else {
229
        t0 = t1 = (vp9_dct_value_tokens_ptr + x)->token;
John Koleszar's avatar
John Koleszar committed
230
      }
231
      if (next < default_eob) {
232
        band = band_translate[i + 1];
233
        if (t0 != EOB_TOKEN) {
234
          pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
235
          rate0 += mb->token_costs[tx_size][type][ref][band][!x][pt]
236
                                  [tokens[next][0].token];
John Koleszar's avatar
John Koleszar committed
237
        }
238
        if (t1 != EOB_TOKEN) {
239
          pt = trellis_get_coeff_context(scan, nb, i, t1, token_cache);
240
          rate1 += mb->token_costs[tx_size][type][ref][band][!x][pt]
241
                                  [tokens[next][1].token];
John Koleszar's avatar
John Koleszar committed
242
243
        }
      }
John Koleszar's avatar
John Koleszar committed
244

245
      UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
246
247
      /* And pick the best. */
      best = rd_cost1 < rd_cost0;
248
      base_bits = *(vp9_dct_value_cost_ptr + x);
John Koleszar's avatar
John Koleszar committed
249
250

      if (shortcut) {
251
        dx -= (dequant_ptr[rc != 0] + sz) ^ sz;
John Koleszar's avatar
John Koleszar committed
252
253
254
255
256
257
258
        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;
259
      best_index[i][1] = best;
John Koleszar's avatar
John Koleszar committed
260
261
      /* Finally, make this the new head of the trellis. */
      next = i;
262
263
264
265
    } 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.
       */
266
      band = band_translate[i + 1];
John Koleszar's avatar
John Koleszar committed
267
268
269
      t0 = tokens[next][0].token;
      t1 = tokens[next][1].token;
      /* Update the cost of each path if we're past the EOB token. */
270
      if (t0 != EOB_TOKEN) {
271
        tokens[next][0].rate +=
272
            mb->token_costs[tx_size][type][ref][band][1][0][t0];
John Koleszar's avatar
John Koleszar committed
273
274
        tokens[next][0].token = ZERO_TOKEN;
      }
275
      if (t1 != EOB_TOKEN) {
276
        tokens[next][1].rate +=
277
            mb->token_costs[tx_size][type][ref][band][1][0][t1];
John Koleszar's avatar
John Koleszar committed
278
279
        tokens[next][1].token = ZERO_TOKEN;
      }
280
      best_index[i][0] = best_index[i][1] = 0;
John Koleszar's avatar
John Koleszar committed
281
      /* Don't update next, because we didn't add a new node. */
282
    }
John Koleszar's avatar
John Koleszar committed
283
284
285
  }

  /* Now pick the best path through the whole trellis. */
286
  band = band_translate[i + 1];
287
  pt = combine_entropy_contexts(*a, *l);
John Koleszar's avatar
John Koleszar committed
288
289
290
291
292
293
  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;
294
295
  rate0 += mb->token_costs[tx_size][type][ref][band][0][pt][t0];
  rate1 += mb->token_costs[tx_size][type][ref][band][0][pt][t1];
296
  UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
297
298
  best = rd_cost1 < rd_cost0;
  final_eob = i0 - 1;
299
300
  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
301
302
  for (i = next; i < eob; i = next) {
    x = tokens[i][best].qc;
303
    if (x) {
John Koleszar's avatar
John Koleszar committed
304
      final_eob = i;
305
    }
306
    rc = scan[i];
John Koleszar's avatar
John Koleszar committed
307
    qcoeff_ptr[rc] = x;
308
    dqcoeff_ptr[rc] = (x * dequant_ptr[rc != 0]) / mul;
309

John Koleszar's avatar
John Koleszar committed
310
    next = tokens[i][best].next;
311
    best = best_index[i][best];
John Koleszar's avatar
John Koleszar committed
312
313
314
  }
  final_eob++;

315
  mb->plane[plane].eobs[block] = final_eob;
316
  *a = *l = (final_eob > 0);
317
318
}

319
void vp9_optimize_b(int plane, int block, BLOCK_SIZE plane_bsize,
320
                    TX_SIZE tx_size, MACROBLOCK *mb, struct optimize_ctx *ctx) {
John Koleszar's avatar
John Koleszar committed
321
  int x, y;
322
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x, &y);
323
324
  optimize_b(mb, plane, block, plane_bsize,
             &ctx->ta[plane][x], &ctx->tl[plane][y], tx_size);
325
326
}

327
static void optimize_init_b(int plane, BLOCK_SIZE bsize,
328
                            struct encode_b_args *args) {
329
330
  const MACROBLOCKD *xd = &args->x->e_mbd;
  const struct macroblockd_plane* const pd = &xd->plane[plane];
331
  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
332
333
  const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
  const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
334
  const MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
335
  const TX_SIZE tx_size = plane ? get_uv_tx_size(mbmi) : mbmi->tx_size;
336

337
338
339
  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);
340
}
341
void vp9_xform_quant(int plane, int block, BLOCK_SIZE plane_bsize,
342
                     TX_SIZE tx_size, void *arg) {
343
344
345
  struct encode_b_args* const args = arg;
  MACROBLOCK* const x = args->x;
  MACROBLOCKD* const xd = &x->e_mbd;
346
347
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
348
  int16_t *coeff = BLOCK_OFFSET(p->coeff, block);
349
  int16_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
350
  int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
351
  const scan_order *scan_order;
352
  uint16_t *eob = &p->eobs[block];
353
354
  const int diff_stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
  int i, j;
355
  int16_t *src_diff;
356
357
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
  src_diff = &p->src_diff[4 * (j * diff_stride + i)];
358
359

  switch (tx_size) {
360
    case TX_32X32:
361
      scan_order = &vp9_default_scan_orders[TX_32X32];
362
      if (x->use_lp32x32fdct)
363
        vp9_fdct32x32_rd(src_diff, coeff, diff_stride);
364
      else
365
        vp9_fdct32x32(src_diff, coeff, diff_stride);
366
367
      vp9_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round,
                           p->quant, p->quant_shift, qcoeff, dqcoeff,
368
369
                           pd->dequant, p->zbin_extra, eob, scan_order->scan,
                           scan_order->iscan);
370
371
      break;
    case TX_16X16:
372
      scan_order = &vp9_default_scan_orders[TX_16X16];
373
      vp9_fdct16x16(src_diff, coeff, diff_stride);
374
375
      vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
                     p->quant, p->quant_shift, qcoeff, dqcoeff,
376
377
                     pd->dequant, p->zbin_extra, eob,
                     scan_order->scan, scan_order->iscan);
378
379
      break;
    case TX_8X8:
380
      scan_order = &vp9_default_scan_orders[TX_8X8];
381
      vp9_fdct8x8(src_diff, coeff, diff_stride);
382
383
      vp9_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round,
                     p->quant, p->quant_shift, qcoeff, dqcoeff,
384
385
                     pd->dequant, p->zbin_extra, eob,
                     scan_order->scan, scan_order->iscan);
386
387
      break;
    case TX_4X4:
388
      scan_order = &vp9_default_scan_orders[TX_4X4];
389
      x->fwd_txm4x4(src_diff, coeff, diff_stride);
390
391
      vp9_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round,
                     p->quant, p->quant_shift, qcoeff, dqcoeff,
392
393
                     pd->dequant, p->zbin_extra, eob,
                     scan_order->scan, scan_order->iscan);
394
395
396
      break;
    default:
      assert(0);
John Koleszar's avatar
John Koleszar committed
397
  }
398
399
}

400
static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize,
401
                         TX_SIZE tx_size, void *arg) {
402
403
404
  struct encode_b_args *const args = arg;
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
405
  struct optimize_ctx *const ctx = args->ctx;
406
  struct macroblock_plane *const p = &x->plane[plane];
407
  struct macroblockd_plane *const pd = &xd->plane[plane];
408
  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
409
410
411
412
  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];
413
414
415
416

  // 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) {
417
    p->eobs[block] = 0;
418
    ctx->ta[plane][i] = 0;
419
    ctx->tl[plane][j] = 0;
420
421
422
    return;
  }

423
  if (!x->skip_recode)
424
    vp9_xform_quant(plane, block, plane_bsize, tx_size, arg);
425

426
  if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
427
    vp9_optimize_b(plane, block, plane_bsize, tx_size, x, ctx);
428
  } else {
429
430
    ctx->ta[plane][i] = p->eobs[block] > 0;
    ctx->tl[plane][j] = p->eobs[block] > 0;
431
  }
John Koleszar's avatar
John Koleszar committed
432

Jim Bankoski's avatar
Jim Bankoski committed
433
434
435
  if (p->eobs[block])
    *(args->skip_coeff) = 0;

436
  if (x->skip_encode || p->eobs[block] == 0)
437
    return;
438

439
  switch (tx_size) {
440
    case TX_32X32:
441
      vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
442
443
      break;
    case TX_16X16:
444
      vp9_idct16x16_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
445
446
      break;
    case TX_8X8:
447
      vp9_idct8x8_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
448
449
      break;
    case TX_4X4:
450
451
452
      // 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.
453
      xd->itxm_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
454
      break;
455
    default:
James Zern's avatar
James Zern committed
456
      assert(0 && "Invalid transform size");
457
  }
John Koleszar's avatar
John Koleszar committed
458
}
459
460
461
462
463
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;
464
  struct macroblock_plane *const p = &x->plane[plane];
465
466
  struct macroblockd_plane *const pd = &xd->plane[plane];
  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
467
468
469
470
  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];
471
472
473

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

474
  if (p->eobs[block] == 0)
475
476
    return;

477
  xd->itxm_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
478
479
}

480
void vp9_encode_sby(MACROBLOCK *x, BLOCK_SIZE bsize) {
481
  MACROBLOCKD *const xd = &x->e_mbd;
482
  struct optimize_ctx ctx;
Jim Bankoski's avatar
Jim Bankoski committed
483
484
  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
  struct encode_b_args arg = {x, &ctx, &mbmi->skip_coeff};
485
486
487

  vp9_subtract_sby(x, bsize);
  if (x->optimize)
488
    optimize_init_b(0, bsize, &arg);
489

490
491
  vp9_foreach_transformed_block_in_plane(xd, bsize, 0, encode_block_pass1,
                                         &arg);
492
493
}

494
void vp9_encode_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
495
  MACROBLOCKD *const xd = &x->e_mbd;
496
  struct optimize_ctx ctx;
Jim Bankoski's avatar
Jim Bankoski committed
497
498
  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
499

500
  if (!x->skip_recode)
501
    vp9_subtract_sb(x, bsize);
502

503
  if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
504
505
506
507
    int i;
    for (i = 0; i < MAX_MB_PLANE; ++i)
      optimize_init_b(i, bsize, &arg);
  }
Scott LaVarnway's avatar
Scott LaVarnway committed
508

509
  vp9_foreach_transformed_block(xd, bsize, encode_block, &arg);
John Koleszar's avatar
John Koleszar committed
510
}
511

512
void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
513
                            TX_SIZE tx_size, void *arg) {
514
  struct encode_b_args* const args = arg;
515
516
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
517
  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
518
519
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
520
  int16_t *coeff = BLOCK_OFFSET(p->coeff, block);
521
  int16_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
522
  int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
523
  const scan_order *scan_order;
524
  TX_TYPE tx_type;
525
  MB_PREDICTION_MODE mode;
526
527
  const int bwl = b_width_log2(plane_bsize);
  const int diff_stride = 4 * (1 << bwl);
528
529
  uint8_t *src, *dst;
  int16_t *src_diff;
530
  uint16_t *eob = &p->eobs[block];
531
532
533
534
535
  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)];
536

537
  // if (x->optimize)
538
  // vp9_optimize_b(plane, block, plane_bsize, tx_size, x, args->ctx);
539

540
  switch (tx_size) {
541
    case TX_32X32:
542
      scan_order = &vp9_default_scan_orders[TX_32X32];
543
      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
544
      vp9_predict_intra_block(xd, block >> 6, bwl, TX_32X32, mode,
545
546
                              x->skip_encode ? src : dst,
                              x->skip_encode ? p->src.stride : pd->dst.stride,
547
                              dst, pd->dst.stride, i, j, plane);
548
      if (!x->skip_recode) {
549
        vp9_subtract_block(32, 32, src_diff, diff_stride,
550
551
                           src, p->src.stride, dst, pd->dst.stride);
        if (x->use_lp32x32fdct)
552
          vp9_fdct32x32_rd(src_diff, coeff, diff_stride);
553
        else
554
          vp9_fdct32x32(src_diff, coeff, diff_stride);
555
556
        vp9_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round,
                             p->quant, p->quant_shift, qcoeff, dqcoeff,
557
558
                             pd->dequant, p->zbin_extra, eob, scan_order->scan,
                             scan_order->iscan);
559
      }
560
      if (!x->skip_encode && *eob)
Yaowu Xu's avatar
Yaowu Xu committed
561
        vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, *eob);
562
563
      break;
    case TX_16X16:
564
      tx_type = get_tx_type_16x16(pd->plane_type, xd);
565
      scan_order = &vp9_scan_orders[TX_16X16][tx_type];
566
      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
567
      vp9_predict_intra_block(xd, block >> 4, bwl, TX_16X16, mode,
568
569
                              x->skip_encode ? src : dst,
                              x->skip_encode ? p->src.stride : pd->dst.stride,
570
                              dst, pd->dst.stride, i, j, plane);
571
      if (!x->skip_recode) {
572
        vp9_subtract_block(16, 16, src_diff, diff_stride,
573
                           src, p->src.stride, dst, pd->dst.stride);
574
        vp9_fht16x16(tx_type, src_diff, coeff, diff_stride);
575
576
        vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
                       p->quant, p->quant_shift, qcoeff, dqcoeff,
577
578
                       pd->dequant, p->zbin_extra, eob, scan_order->scan,
                       scan_order->iscan);
579
      }
580
      if (!x->skip_encode && *eob)
581
        vp9_iht16x16_add(tx_type, dqcoeff, dst, pd->dst.stride, *eob);
582
583
      break;
    case TX_8X8:
584
      tx_type = get_tx_type_8x8(pd->plane_type, xd);
585
      scan_order = &vp9_scan_orders[TX_8X8][tx_type];
586
      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
587
      vp9_predict_intra_block(xd, block >> 2, bwl, TX_8X8, mode,
588
589
                              x->skip_encode ? src : dst,
                              x->skip_encode ? p->src.stride : pd->dst.stride,
590
                              dst, pd->dst.stride, i, j, plane);
591
      if (!x->skip_recode) {
592
        vp9_subtract_block(8, 8, src_diff, diff_stride,
593
                           src, p->src.stride, dst, pd->dst.stride);
594
        vp9_fht8x8(tx_type, src_diff, coeff, diff_stride);
595
596
        vp9_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round, p->quant,
                       p->quant_shift, qcoeff, dqcoeff,
597
598
                       pd->dequant, p->zbin_extra, eob, scan_order->scan,
                       scan_order->iscan);
599
      }
600
      if (!x->skip_encode && *eob)
601
        vp9_iht8x8_add(tx_type, dqcoeff, dst, pd->dst.stride, *eob);
602
603
      break;
    case TX_4X4:
604
      tx_type = get_tx_type_4x4(pd->plane_type, xd, block);
605
      scan_order = &vp9_scan_orders[TX_4X4][tx_type];
606
      if (mbmi->sb_type < BLOCK_8X8 && plane == 0)
607
        mode = xd->mi_8x8[0]->bmi[block].as_mode;
608
      else
609
        mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
610

611
      vp9_predict_intra_block(xd, block, bwl, TX_4X4, mode,
612
613
                              x->skip_encode ? src : dst,
                              x->skip_encode ? p->src.stride : pd->dst.stride,
614
                              dst, pd->dst.stride, i, j, plane);
615
616

      if (!x->skip_recode) {
617
        vp9_subtract_block(4, 4, src_diff, diff_stride,
618
619
                           src, p->src.stride, dst, pd->dst.stride);
        if (tx_type != DCT_DCT)
620
          vp9_short_fht4x4(src_diff, coeff, diff_stride, tx_type);
621
        else
622
          x->fwd_txm4x4(src_diff, coeff, diff_stride);
623
624
        vp9_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round, p->quant,
                       p->quant_shift, qcoeff, dqcoeff,
625
626
                       pd->dequant, p->zbin_extra, eob, scan_order->scan,
                       scan_order->iscan);
627
628
      }

629
      if (!x->skip_encode && *eob) {
630
631
632
633
        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.
634
          xd->itxm_add(dqcoeff, dst, pd->dst.stride, *eob);
635
        else
636
          vp9_iht4x4_16_add(dqcoeff, dst, pd->dst.stride, tx_type);
637
      }
638
      break;
639
640
    default:
      assert(0);
641
  }
Jim Bankoski's avatar
Jim Bankoski committed
642
643
  if (*eob)
    *(args->skip_coeff) = 0;
644
645
}

646
void vp9_encode_intra_block_y(MACROBLOCK *x, BLOCK_SIZE bsize) {
647
648
  MACROBLOCKD* const xd = &x->e_mbd;
  struct optimize_ctx ctx;
Jim Bankoski's avatar
Jim Bankoski committed
649
650
  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
  struct encode_b_args arg = {x, &ctx, &mbmi->skip_coeff};
651

652
  vp9_foreach_transformed_block_in_plane(xd, bsize, 0, vp9_encode_block_intra,
653
                                     &arg);
654
}
655
void vp9_encode_intra_block_uv(MACROBLOCK *x, BLOCK_SIZE bsize) {
656
657
  MACROBLOCKD* const xd = &x->e_mbd;
  struct optimize_ctx ctx;
Jim Bankoski's avatar
Jim Bankoski committed
658
659
  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
  struct encode_b_args arg = {x, &ctx, &mbmi->skip_coeff};
660
  vp9_foreach_transformed_block_uv(xd, bsize, vp9_encode_block_intra, &arg);
661
}
662

663
664
665
666
667
668
669
670
671
672
673
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);
}