vp9_encodemb.c 25 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
void vp9_setup_interp_filters(MACROBLOCKD *xd, INTERP_FILTER filter,
29
30
31
32
                              VP9_COMMON *cm) {
  if (xd->mi_8x8 && xd->mi_8x8[0]) {
    MB_MODE_INFO *const mbmi = &xd->mi_8x8[0]->mbmi;

33
34
    set_ref_ptrs(cm, xd, mbmi->ref_frame[0] - LAST_FRAME,
                 mbmi->ref_frame[1] - LAST_FRAME);
35

36
  } else {
37
    set_ref_ptrs(cm, xd, -1, -1);
38
39
  }

40
41
42
  xd->interp_kernel = vp9_get_interp_kernel(filter == SWITCHABLE ? EIGHTTAP
                                                                 : filter);
  assert(((intptr_t)xd->interp_kernel & 0xff) == 0);
43
44
}

45
46
47
48
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
49
  int r, c;
John Koleszar's avatar
John Koleszar committed
50

51
52
  for (r = 0; r < rows; r++) {
    for (c = 0; c < cols; c++)
John Koleszar's avatar
John Koleszar committed
53
54
      diff_ptr[c] = src_ptr[c] - pred_ptr[c];

55
56
    diff_ptr += diff_stride;
    pred_ptr += pred_stride;
John Koleszar's avatar
John Koleszar committed
57
58
    src_ptr  += src_stride;
  }
John Koleszar's avatar
John Koleszar committed
59
60
}

61
static void subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
62
  struct macroblock_plane *const p = &x->plane[plane];
63
64
65
66
  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];
67

68
  vp9_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
69
                     pd->dst.buf, pd->dst.stride);
Yaowu Xu's avatar
Yaowu Xu committed
70
71
}

72
void vp9_subtract_sby(MACROBLOCK *x, BLOCK_SIZE bsize) {
73
  subtract_plane(x, bsize, 0);
74
75
}

76
void vp9_subtract_sbuv(MACROBLOCK *x, BLOCK_SIZE bsize) {
77
  int i;
78

79
80
  for (i = 1; i < MAX_MB_PLANE; i++)
    subtract_plane(x, bsize, i);
81
82
}

83
void vp9_subtract_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
84
85
  vp9_subtract_sby(x, bsize);
  vp9_subtract_sbuv(x, bsize);
John Koleszar's avatar
John Koleszar committed
86
87
}

88
#define RDTRUNC(RM, DM, R, D) ((128 + (R) * (RM)) & 0xFF)
89
typedef struct vp9_token_state vp9_token_state;
90

91
struct vp9_token_state {
92
93
  int           rate;
  int           error;
Daniel Kang's avatar
Daniel Kang committed
94
  int           next;
95
96
97
98
  signed char   token;
  short         qc;
};

99
// TODO(jimbankoski): experiment to find optimal RD numbers.
100
101
#define Y1_RD_MULT 4
#define UV_RD_MULT 2
102

John Koleszar's avatar
John Koleszar committed
103
104
105
static const int plane_rd_mult[4] = {
  Y1_RD_MULT,
  UV_RD_MULT,
106
107
};

108
109
110
111
112
113
114
115
116
117
#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);\
  }\
}

118
119
// This function is a place holder for now but may ultimately need
// to scan previous tokens to work out the correct context.
120
121
static int trellis_get_coeff_context(const int16_t *scan,
                                     const int16_t *nb,
122
                                     int idx, int token,
123
                                     uint8_t *token_cache) {
124
  int bak = token_cache[scan[idx]], pt;
125
  token_cache[scan[idx]] = vp9_pt_energy_class[token];
126
  pt = get_coef_context(nb, token_cache, idx + 1);
127
  token_cache[scan[idx]] = bak;
128
  return pt;
129
130
}

131
static void optimize_b(MACROBLOCK *mb,
132
                       int plane, int block, BLOCK_SIZE plane_bsize,
133
                       ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
John Koleszar's avatar
John Koleszar committed
134
                       TX_SIZE tx_size) {
135
  MACROBLOCKD *const xd = &mb->e_mbd;
136
  struct macroblock_plane *p = &mb->plane[plane];
137
  struct macroblockd_plane *pd = &xd->plane[plane];
138
  const int ref = is_inter_block(&xd->mi_8x8[0]->mbmi);
139
140
  vp9_token_state tokens[1025][2];
  unsigned best_index[1025][2];
141
  const int16_t *coeff_ptr = BLOCK_OFFSET(mb->plane[plane].coeff, block);
142
143
  int16_t *qcoeff_ptr;
  int16_t *dqcoeff_ptr;
144
  int eob = p->eobs[block], final_eob, sz = 0;
145
  const int i0 = 0;
146
  int rc, x, next, i;
147
148
149
  int64_t rdmult, rddiv, rd_cost0, rd_cost1;
  int rate0, rate1, error0, error1, t0, t1;
  int best, band, pt;
150
  PLANE_TYPE type = pd->plane_type;
John Koleszar's avatar
John Koleszar committed
151
  int err_mult = plane_rd_mult[type];
152
  const int default_eob = 16 << (tx_size << 1);
153

154
  const int mul = 1 + (tx_size == TX_32X32);
155
  uint8_t token_cache[1024];
156
  const int16_t *dequant_ptr = pd->dequant;
157
  const uint8_t *const band_translate = get_band_translate(tx_size);
158
159
160
  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
161

John Koleszar's avatar
John Koleszar committed
162
  assert((!type && !plane) || (type && plane));
163
  dqcoeff_ptr = BLOCK_OFFSET(pd->dqcoeff, block);
164
  qcoeff_ptr = BLOCK_OFFSET(p->qcoeff, block);
John Koleszar's avatar
John Koleszar committed
165
  assert(eob <= default_eob);
John Koleszar's avatar
John Koleszar committed
166
167
168

  /* Now set up a Viterbi trellis to evaluate alternative roundings. */
  rdmult = mb->rdmult * err_mult;
169
  if (!is_inter_block(&mb->e_mbd.mi_8x8[0]->mbmi))
John Koleszar's avatar
John Koleszar committed
170
171
172
173
174
    rdmult = (rdmult * 9) >> 4;
  rddiv = mb->rddiv;
  /* Initialize the sentinel node of the trellis. */
  tokens[eob][0].rate = 0;
  tokens[eob][0].error = 0;
175
  tokens[eob][0].next = default_eob;
176
  tokens[eob][0].token = EOB_TOKEN;
John Koleszar's avatar
John Koleszar committed
177
178
179
  tokens[eob][0].qc = 0;
  *(tokens[eob] + 1) = *(tokens[eob] + 0);
  next = eob;
180
  for (i = 0; i < eob; i++)
181
182
    token_cache[scan[i]] = vp9_pt_energy_class[vp9_dct_value_tokens_ptr[
        qcoeff_ptr[scan[i]]].token];
183

John Koleszar's avatar
John Koleszar committed
184
  for (i = eob; i-- > i0;) {
185
    int base_bits, d2, dx;
John Koleszar's avatar
John Koleszar committed
186

187
    rc = scan[i];
John Koleszar's avatar
John Koleszar committed
188
189
190
191
192
193
194
195
196
    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;
197
      t0 = (vp9_dct_value_tokens_ptr + x)->token;
John Koleszar's avatar
John Koleszar committed
198
      /* Consider both possible successor states. */
199
      if (next < default_eob) {
200
        band = band_translate[i + 1];
201
        pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
John Koleszar's avatar
John Koleszar committed
202
        rate0 +=
203
          mb->token_costs[tx_size][type][ref][band][0][pt]
204
                         [tokens[next][0].token];
John Koleszar's avatar
John Koleszar committed
205
        rate1 +=
206
          mb->token_costs[tx_size][type][ref][band][0][pt]
207
                         [tokens[next][1].token];
John Koleszar's avatar
John Koleszar committed
208
      }
209
      UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
210
211
      /* And pick the best. */
      best = rd_cost1 < rd_cost0;
212
      base_bits = *(vp9_dct_value_cost_ptr + x);
213
      dx = mul * (dqcoeff_ptr[rc] - coeff_ptr[rc]);
John Koleszar's avatar
John Koleszar committed
214
215
216
217
218
219
      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;
220
      best_index[i][0] = best;
221

John Koleszar's avatar
John Koleszar committed
222
223
224
225
      /* Evaluate the second possibility for this state. */
      rate0 = tokens[next][0].rate;
      rate1 = tokens[next][1].rate;

226
227
228
      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
229
230
231
232
233
234
235
236
237
238
239
240
241
        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.
242
         */
243
244
        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
245
      } else {
246
        t0 = t1 = (vp9_dct_value_tokens_ptr + x)->token;
John Koleszar's avatar
John Koleszar committed
247
      }
248
      if (next < default_eob) {
249
        band = band_translate[i + 1];
250
        if (t0 != EOB_TOKEN) {
251
          pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
252
          rate0 += mb->token_costs[tx_size][type][ref][band][!x][pt]
253
                                  [tokens[next][0].token];
John Koleszar's avatar
John Koleszar committed
254
        }
255
        if (t1 != EOB_TOKEN) {
256
          pt = trellis_get_coeff_context(scan, nb, i, t1, token_cache);
257
          rate1 += mb->token_costs[tx_size][type][ref][band][!x][pt]
258
                                  [tokens[next][1].token];
John Koleszar's avatar
John Koleszar committed
259
260
        }
      }
John Koleszar's avatar
John Koleszar committed
261

262
      UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
263
264
      /* And pick the best. */
      best = rd_cost1 < rd_cost0;
265
      base_bits = *(vp9_dct_value_cost_ptr + x);
John Koleszar's avatar
John Koleszar committed
266
267

      if (shortcut) {
268
        dx -= (dequant_ptr[rc != 0] + sz) ^ sz;
John Koleszar's avatar
John Koleszar committed
269
270
271
272
273
274
275
        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;
276
      best_index[i][1] = best;
John Koleszar's avatar
John Koleszar committed
277
278
      /* Finally, make this the new head of the trellis. */
      next = i;
279
280
281
282
    } 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.
       */
283
      band = band_translate[i + 1];
John Koleszar's avatar
John Koleszar committed
284
285
286
      t0 = tokens[next][0].token;
      t1 = tokens[next][1].token;
      /* Update the cost of each path if we're past the EOB token. */
287
      if (t0 != EOB_TOKEN) {
288
        tokens[next][0].rate +=
289
            mb->token_costs[tx_size][type][ref][band][1][0][t0];
John Koleszar's avatar
John Koleszar committed
290
291
        tokens[next][0].token = ZERO_TOKEN;
      }
292
      if (t1 != EOB_TOKEN) {
293
        tokens[next][1].rate +=
294
            mb->token_costs[tx_size][type][ref][band][1][0][t1];
John Koleszar's avatar
John Koleszar committed
295
296
        tokens[next][1].token = ZERO_TOKEN;
      }
297
      best_index[i][0] = best_index[i][1] = 0;
John Koleszar's avatar
John Koleszar committed
298
      /* Don't update next, because we didn't add a new node. */
299
    }
John Koleszar's avatar
John Koleszar committed
300
301
302
  }

  /* Now pick the best path through the whole trellis. */
303
  band = band_translate[i + 1];
304
  pt = combine_entropy_contexts(*a, *l);
John Koleszar's avatar
John Koleszar committed
305
306
307
308
309
310
  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;
311
312
  rate0 += mb->token_costs[tx_size][type][ref][band][0][pt][t0];
  rate1 += mb->token_costs[tx_size][type][ref][band][0][pt][t1];
313
  UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
314
315
  best = rd_cost1 < rd_cost0;
  final_eob = i0 - 1;
316
317
  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
318
319
  for (i = next; i < eob; i = next) {
    x = tokens[i][best].qc;
320
    if (x) {
John Koleszar's avatar
John Koleszar committed
321
      final_eob = i;
322
    }
323
    rc = scan[i];
John Koleszar's avatar
John Koleszar committed
324
    qcoeff_ptr[rc] = x;
325
    dqcoeff_ptr[rc] = (x * dequant_ptr[rc != 0]) / mul;
326

John Koleszar's avatar
John Koleszar committed
327
    next = tokens[i][best].next;
328
    best = best_index[i][best];
John Koleszar's avatar
John Koleszar committed
329
330
331
  }
  final_eob++;

332
  mb->plane[plane].eobs[block] = final_eob;
333
  *a = *l = (final_eob > 0);
334
335
}

336
void vp9_optimize_b(int plane, int block, BLOCK_SIZE plane_bsize,
337
                    TX_SIZE tx_size, MACROBLOCK *mb, struct optimize_ctx *ctx) {
John Koleszar's avatar
John Koleszar committed
338
  int x, y;
339
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x, &y);
340
341
  optimize_b(mb, plane, block, plane_bsize,
             &ctx->ta[plane][x], &ctx->tl[plane][y], tx_size);
342
343
}

344
static void optimize_init_b(int plane, BLOCK_SIZE bsize,
345
                            struct encode_b_args *args) {
346
347
  const MACROBLOCKD *xd = &args->x->e_mbd;
  const struct macroblockd_plane* const pd = &xd->plane[plane];
348
  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
349
350
  const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
  const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
351
  const MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
352
  const TX_SIZE tx_size = plane ? get_uv_tx_size(mbmi) : mbmi->tx_size;
353

354
355
356
  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);
357
}
358
void vp9_xform_quant(int plane, int block, BLOCK_SIZE plane_bsize,
359
                     TX_SIZE tx_size, void *arg) {
360
361
362
  struct encode_b_args* const args = arg;
  MACROBLOCK* const x = args->x;
  MACROBLOCKD* const xd = &x->e_mbd;
363
364
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
365
  int16_t *coeff = BLOCK_OFFSET(p->coeff, block);
366
  int16_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
367
  int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
368
  const scan_order *scan_order;
369
  uint16_t *eob = &p->eobs[block];
370
371
  const int diff_stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
  int i, j;
372
  int16_t *src_diff;
373
374
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
  src_diff = &p->src_diff[4 * (j * diff_stride + i)];
375
376

  switch (tx_size) {
377
    case TX_32X32:
378
      scan_order = &vp9_default_scan_orders[TX_32X32];
379
      if (x->use_lp32x32fdct)
380
        vp9_fdct32x32_rd(src_diff, coeff, diff_stride);
381
      else
382
        vp9_fdct32x32(src_diff, coeff, diff_stride);
383
384
      vp9_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round,
                           p->quant, p->quant_shift, qcoeff, dqcoeff,
385
386
                           pd->dequant, p->zbin_extra, eob, scan_order->scan,
                           scan_order->iscan);
387
388
      break;
    case TX_16X16:
389
      scan_order = &vp9_default_scan_orders[TX_16X16];
390
      vp9_fdct16x16(src_diff, coeff, diff_stride);
391
392
      vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
                     p->quant, p->quant_shift, qcoeff, dqcoeff,
393
394
                     pd->dequant, p->zbin_extra, eob,
                     scan_order->scan, scan_order->iscan);
395
396
      break;
    case TX_8X8:
397
      scan_order = &vp9_default_scan_orders[TX_8X8];
398
      vp9_fdct8x8(src_diff, coeff, diff_stride);
399
400
      vp9_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round,
                     p->quant, p->quant_shift, qcoeff, dqcoeff,
401
402
                     pd->dequant, p->zbin_extra, eob,
                     scan_order->scan, scan_order->iscan);
403
404
      break;
    case TX_4X4:
405
      scan_order = &vp9_default_scan_orders[TX_4X4];
406
      x->fwd_txm4x4(src_diff, coeff, diff_stride);
407
408
      vp9_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round,
                     p->quant, p->quant_shift, qcoeff, dqcoeff,
409
410
                     pd->dequant, p->zbin_extra, eob,
                     scan_order->scan, scan_order->iscan);
411
412
413
      break;
    default:
      assert(0);
John Koleszar's avatar
John Koleszar committed
414
  }
415
416
}

417
static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize,
418
                         TX_SIZE tx_size, void *arg) {
419
420
421
  struct encode_b_args *const args = arg;
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
422
  struct optimize_ctx *const ctx = args->ctx;
423
  struct macroblock_plane *const p = &x->plane[plane];
424
  struct macroblockd_plane *const pd = &xd->plane[plane];
425
  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
426
427
428
429
  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];
430
431
432
433

  // 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) {
434
    p->eobs[block] = 0;
435
    ctx->ta[plane][i] = 0;
436
    ctx->tl[plane][j] = 0;
437
438
439
    return;
  }

440
  if (!x->skip_recode)
441
    vp9_xform_quant(plane, block, plane_bsize, tx_size, arg);
442

443
  if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
444
    vp9_optimize_b(plane, block, plane_bsize, tx_size, x, ctx);
445
  } else {
446
447
    ctx->ta[plane][i] = p->eobs[block] > 0;
    ctx->tl[plane][j] = p->eobs[block] > 0;
448
  }
John Koleszar's avatar
John Koleszar committed
449

Jim Bankoski's avatar
Jim Bankoski committed
450
451
452
  if (p->eobs[block])
    *(args->skip_coeff) = 0;

453
  if (x->skip_encode || p->eobs[block] == 0)
454
    return;
455

456
  switch (tx_size) {
457
    case TX_32X32:
458
      vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
459
460
      break;
    case TX_16X16:
461
      vp9_idct16x16_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
462
463
      break;
    case TX_8X8:
464
      vp9_idct8x8_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
465
466
      break;
    case TX_4X4:
467
468
469
      // 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.
470
      xd->itxm_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
471
      break;
472
    default:
James Zern's avatar
James Zern committed
473
      assert(0 && "Invalid transform size");
474
  }
John Koleszar's avatar
John Koleszar committed
475
}
476
477
478
479
480
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;
481
  struct macroblock_plane *const p = &x->plane[plane];
482
483
  struct macroblockd_plane *const pd = &xd->plane[plane];
  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
484
485
486
487
  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];
488
489
490

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

491
  if (p->eobs[block] == 0)
492
493
    return;

494
  xd->itxm_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
495
496
}

497
void vp9_encode_sby(MACROBLOCK *x, BLOCK_SIZE bsize) {
498
  MACROBLOCKD *const xd = &x->e_mbd;
499
  struct optimize_ctx ctx;
Jim Bankoski's avatar
Jim Bankoski committed
500
501
  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
  struct encode_b_args arg = {x, &ctx, &mbmi->skip_coeff};
502
503
504

  vp9_subtract_sby(x, bsize);
  if (x->optimize)
505
    optimize_init_b(0, bsize, &arg);
506

507
  foreach_transformed_block_in_plane(xd, bsize, 0, encode_block_pass1, &arg);
508
509
}

510
void vp9_encode_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
511
  MACROBLOCKD *const xd = &x->e_mbd;
512
  struct optimize_ctx ctx;
Jim Bankoski's avatar
Jim Bankoski committed
513
514
  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
515

516
  if (!x->skip_recode)
517
    vp9_subtract_sb(x, bsize);
518

519
  if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
520
521
522
523
    int i;
    for (i = 0; i < MAX_MB_PLANE; ++i)
      optimize_init_b(i, bsize, &arg);
  }
Scott LaVarnway's avatar
Scott LaVarnway committed
524

Scott LaVarnway's avatar
Scott LaVarnway committed
525
  foreach_transformed_block(xd, bsize, encode_block, &arg);
John Koleszar's avatar
John Koleszar committed
526
}
527

528
void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
529
                            TX_SIZE tx_size, void *arg) {
530
  struct encode_b_args* const args = arg;
531
532
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
533
  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
534
535
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
536
  int16_t *coeff = BLOCK_OFFSET(p->coeff, block);
537
  int16_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
538
  int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
539
  const scan_order *scan_order;
540
  TX_TYPE tx_type;
541
  MB_PREDICTION_MODE mode;
542
543
  const int bwl = b_width_log2(plane_bsize);
  const int diff_stride = 4 * (1 << bwl);
544
545
  uint8_t *src, *dst;
  int16_t *src_diff;
546
  uint16_t *eob = &p->eobs[block];
547
548
549
550
551
  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)];
552

553
  // if (x->optimize)
554
  // vp9_optimize_b(plane, block, plane_bsize, tx_size, x, args->ctx);
555

556
  switch (tx_size) {
557
    case TX_32X32:
558
      scan_order = &vp9_default_scan_orders[TX_32X32];
559
      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
560
      vp9_predict_intra_block(xd, block >> 6, bwl, TX_32X32, mode,
561
562
                              x->skip_encode ? src : dst,
                              x->skip_encode ? p->src.stride : pd->dst.stride,
563
                              dst, pd->dst.stride, i, j, plane);
564
      if (!x->skip_recode) {
565
        vp9_subtract_block(32, 32, src_diff, diff_stride,
566
567
                           src, p->src.stride, dst, pd->dst.stride);
        if (x->use_lp32x32fdct)
568
          vp9_fdct32x32_rd(src_diff, coeff, diff_stride);
569
        else
570
          vp9_fdct32x32(src_diff, coeff, diff_stride);
571
572
        vp9_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round,
                             p->quant, p->quant_shift, qcoeff, dqcoeff,
573
574
                             pd->dequant, p->zbin_extra, eob, scan_order->scan,
                             scan_order->iscan);
575
      }
576
      if (!x->skip_encode && *eob)
Yaowu Xu's avatar
Yaowu Xu committed
577
        vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, *eob);
578
579
      break;
    case TX_16X16:
580
      tx_type = get_tx_type_16x16(pd->plane_type, xd);
581
      scan_order = &vp9_scan_orders[TX_16X16][tx_type];
582
      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
583
      vp9_predict_intra_block(xd, block >> 4, bwl, TX_16X16, mode,
584
585
                              x->skip_encode ? src : dst,
                              x->skip_encode ? p->src.stride : pd->dst.stride,
586
                              dst, pd->dst.stride, i, j, plane);
587
      if (!x->skip_recode) {
588
        vp9_subtract_block(16, 16, src_diff, diff_stride,
589
                           src, p->src.stride, dst, pd->dst.stride);
590
        vp9_fht16x16(tx_type, src_diff, coeff, diff_stride);
591
592
        vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
                       p->quant, p->quant_shift, qcoeff, dqcoeff,
593
594
                       pd->dequant, p->zbin_extra, eob, scan_order->scan,
                       scan_order->iscan);
595
      }
596
      if (!x->skip_encode && *eob)
597
        vp9_iht16x16_add(tx_type, dqcoeff, dst, pd->dst.stride, *eob);
598
599
      break;
    case TX_8X8:
600
      tx_type = get_tx_type_8x8(pd->plane_type, xd);
601
      scan_order = &vp9_scan_orders[TX_8X8][tx_type];
602
      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
603
      vp9_predict_intra_block(xd, block >> 2, bwl, TX_8X8, mode,
604
605
                              x->skip_encode ? src : dst,
                              x->skip_encode ? p->src.stride : pd->dst.stride,
606
                              dst, pd->dst.stride, i, j, plane);
607
      if (!x->skip_recode) {
608
        vp9_subtract_block(8, 8, src_diff, diff_stride,
609
                           src, p->src.stride, dst, pd->dst.stride);
610
        vp9_fht8x8(tx_type, src_diff, coeff, diff_stride);
611
612
        vp9_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round, p->quant,
                       p->quant_shift, qcoeff, dqcoeff,
613
614
                       pd->dequant, p->zbin_extra, eob, scan_order->scan,
                       scan_order->iscan);
615
      }
616
      if (!x->skip_encode && *eob)
617
        vp9_iht8x8_add(tx_type, dqcoeff, dst, pd->dst.stride, *eob);
618
619
      break;
    case TX_4X4:
620
      tx_type = get_tx_type_4x4(pd->plane_type, xd, block);
621
      scan_order = &vp9_scan_orders[TX_4X4][tx_type];
622
      if (mbmi->sb_type < BLOCK_8X8 && plane == 0)
623
        mode = xd->mi_8x8[0]->bmi[block].as_mode;
624
      else
625
        mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
626

627
      vp9_predict_intra_block(xd, block, bwl, TX_4X4, mode,
628
629
                              x->skip_encode ? src : dst,
                              x->skip_encode ? p->src.stride : pd->dst.stride,
630
                              dst, pd->dst.stride, i, j, plane);
631
632

      if (!x->skip_recode) {
633
        vp9_subtract_block(4, 4, src_diff, diff_stride,
634
635
                           src, p->src.stride, dst, pd->dst.stride);
        if (tx_type != DCT_DCT)
636
          vp9_short_fht4x4(src_diff, coeff, diff_stride, tx_type);
637
        else
638
          x->fwd_txm4x4(src_diff, coeff, diff_stride);
639
640
        vp9_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round, p->quant,
                       p->quant_shift, qcoeff, dqcoeff,
641
642
                       pd->dequant, p->zbin_extra, eob, scan_order->scan,
                       scan_order->iscan);
643
644
      }

645
      if (!x->skip_encode && *eob) {
646
647
648
649
        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.
650
          xd->itxm_add(dqcoeff, dst, pd->dst.stride, *eob);
651
        else
652
          vp9_iht4x4_16_add(dqcoeff, dst, pd->dst.stride, tx_type);
653
      }
654
      break;
655
656
    default:
      assert(0);
657
  }
Jim Bankoski's avatar
Jim Bankoski committed
658
659
  if (*eob)
    *(args->skip_coeff) = 0;
660
661
}

662
void vp9_encode_intra_block_y(MACROBLOCK *x, BLOCK_SIZE bsize) {
663
664
  MACROBLOCKD* const xd = &x->e_mbd;
  struct optimize_ctx ctx;
Jim Bankoski's avatar
Jim Bankoski committed
665
666
  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
  struct encode_b_args arg = {x, &ctx, &mbmi->skip_coeff};
667

668
669
  foreach_transformed_block_in_plane(xd, bsize, 0, vp9_encode_block_intra,
                                     &arg);
670
}
671
void vp9_encode_intra_block_uv(MACROBLOCK *x, BLOCK_SIZE bsize) {
672
673
  MACROBLOCKD* const xd = &x->e_mbd;
  struct optimize_ctx ctx;
Jim Bankoski's avatar
Jim Bankoski committed
674
675
  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
  struct encode_b_args arg = {x, &ctx, &mbmi->skip_coeff};
676
  foreach_transformed_block_uv(xd, bsize, vp9_encode_block_intra, &arg);
677
}
678

679
680
681
682
683
684
685
686
687
688
689
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);
}