vp9_encodemb.c 24.5 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
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
void vp9_setup_interp_filters(MACROBLOCKD *xd,
                              INTERPOLATION_TYPE mcomp_filter_type,
                              VP9_COMMON *cm) {
  if (xd->mi_8x8 && xd->mi_8x8[0]) {
    MB_MODE_INFO *const mbmi = &xd->mi_8x8[0]->mbmi;

    set_scale_factors(xd, mbmi->ref_frame[0] - LAST_FRAME,
                          mbmi->ref_frame[1] - LAST_FRAME,
                          cm->active_ref_scale);
  } else {
    set_scale_factors(xd, -1, -1, cm->active_ref_scale);
  }

  xd->subpix.filter_x = xd->subpix.filter_y =
      vp9_get_filter_kernel(mcomp_filter_type == SWITCHABLE ?
                               EIGHTTAP : mcomp_filter_type);

  assert(((intptr_t)xd->subpix.filter_x & 0xff) == 0);
}

48
49
50
51
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
52
  int r, c;
John Koleszar's avatar
John Koleszar committed
53

54
55
  for (r = 0; r < rows; r++) {
    for (c = 0; c < cols; c++)
John Koleszar's avatar
John Koleszar committed
56
57
      diff_ptr[c] = src_ptr[c] - pred_ptr[c];

58
59
    diff_ptr += diff_stride;
    pred_ptr += pred_stride;
John Koleszar's avatar
John Koleszar committed
60
61
    src_ptr  += src_stride;
  }
John Koleszar's avatar
John Koleszar committed
62
63
}

64
static void subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
65
  struct macroblock_plane *const p = &x->plane[plane];
66
67
68
69
  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];
70

71
  vp9_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
72
                     pd->dst.buf, pd->dst.stride);
Yaowu Xu's avatar
Yaowu Xu committed
73
74
}

75
void vp9_subtract_sby(MACROBLOCK *x, BLOCK_SIZE bsize) {
76
  subtract_plane(x, bsize, 0);
77
78
}

79
void vp9_subtract_sbuv(MACROBLOCK *x, BLOCK_SIZE bsize) {
80
  int i;
81

82
83
  for (i = 1; i < MAX_MB_PLANE; i++)
    subtract_plane(x, bsize, i);
84
85
}

86
void vp9_subtract_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
87
88
  vp9_subtract_sby(x, bsize);
  vp9_subtract_sbuv(x, bsize);
John Koleszar's avatar
John Koleszar committed
89
90
}

91
#define RDTRUNC(RM, DM, R, D) ((128 + (R) * (RM)) & 0xFF)
92
typedef struct vp9_token_state vp9_token_state;
93

94
struct vp9_token_state {
95
96
  int           rate;
  int           error;
Daniel Kang's avatar
Daniel Kang committed
97
  int           next;
98
99
100
101
  signed char   token;
  short         qc;
};

102
// TODO(jimbankoski): experiment to find optimal RD numbers.
103
104
#define Y1_RD_MULT 4
#define UV_RD_MULT 2
105

John Koleszar's avatar
John Koleszar committed
106
107
108
static const int plane_rd_mult[4] = {
  Y1_RD_MULT,
  UV_RD_MULT,
109
110
};

111
112
113
114
115
116
117
118
119
120
#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);\
  }\
}

121
122
// This function is a place holder for now but may ultimately need
// to scan previous tokens to work out the correct context.
123
124
static int trellis_get_coeff_context(const int16_t *scan,
                                     const int16_t *nb,
125
                                     int idx, int token,
126
                                     uint8_t *token_cache) {
127
  int bak = token_cache[scan[idx]], pt;
128
  token_cache[scan[idx]] = vp9_pt_energy_class[token];
129
  pt = get_coef_context(nb, token_cache, idx + 1);
130
  token_cache[scan[idx]] = bak;
131
  return pt;
132
133
}

134
static void optimize_b(MACROBLOCK *mb,
135
                       int plane, int block, BLOCK_SIZE plane_bsize,
136
                       ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
John Koleszar's avatar
John Koleszar committed
137
                       TX_SIZE tx_size) {
138
  MACROBLOCKD *const xd = &mb->e_mbd;
139
  struct macroblock_plane *p = &mb->plane[plane];
140
  struct macroblockd_plane *pd = &xd->plane[plane];
141
  const int ref = is_inter_block(&xd->mi_8x8[0]->mbmi);
142
143
  vp9_token_state tokens[1025][2];
  unsigned best_index[1025][2];
144
  const int16_t *coeff_ptr = BLOCK_OFFSET(mb->plane[plane].coeff, block);
145
146
  int16_t *qcoeff_ptr;
  int16_t *dqcoeff_ptr;
147
  int eob = pd->eobs[block], final_eob, sz = 0;
148
  const int i0 = 0;
149
  int rc, x, next, i;
150
151
152
  int64_t rdmult, rddiv, rd_cost0, rd_cost1;
  int rate0, rate1, error0, error1, t0, t1;
  int best, band, pt;
153
  PLANE_TYPE type = pd->plane_type;
John Koleszar's avatar
John Koleszar committed
154
  int err_mult = plane_rd_mult[type];
155
  const int default_eob = 16 << (tx_size << 1);
156

157
  const int mul = 1 + (tx_size == TX_32X32);
158
  uint8_t token_cache[1024];
159
  const int16_t *dequant_ptr = pd->dequant;
160
  const uint8_t *const band_translate = get_band_translate(tx_size);
161
162
163
  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
164

John Koleszar's avatar
John Koleszar committed
165
  assert((!type && !plane) || (type && plane));
166
  dqcoeff_ptr = BLOCK_OFFSET(pd->dqcoeff, block);
167
  qcoeff_ptr = BLOCK_OFFSET(p->qcoeff, block);
John Koleszar's avatar
John Koleszar committed
168
  assert(eob <= default_eob);
John Koleszar's avatar
John Koleszar committed
169
170
171

  /* Now set up a Viterbi trellis to evaluate alternative roundings. */
  rdmult = mb->rdmult * err_mult;
172
  if (!is_inter_block(&mb->e_mbd.mi_8x8[0]->mbmi))
John Koleszar's avatar
John Koleszar committed
173
174
175
176
177
    rdmult = (rdmult * 9) >> 4;
  rddiv = mb->rddiv;
  /* Initialize the sentinel node of the trellis. */
  tokens[eob][0].rate = 0;
  tokens[eob][0].error = 0;
178
  tokens[eob][0].next = default_eob;
John Koleszar's avatar
John Koleszar committed
179
180
181
182
  tokens[eob][0].token = DCT_EOB_TOKEN;
  tokens[eob][0].qc = 0;
  *(tokens[eob] + 1) = *(tokens[eob] + 0);
  next = eob;
183
  for (i = 0; i < eob; i++)
184
185
    token_cache[scan[i]] = vp9_pt_energy_class[vp9_dct_value_tokens_ptr[
        qcoeff_ptr[scan[i]]].token];
186

John Koleszar's avatar
John Koleszar committed
187
  for (i = eob; i-- > i0;) {
188
    int base_bits, d2, dx;
John Koleszar's avatar
John Koleszar committed
189

190
    rc = scan[i];
John Koleszar's avatar
John Koleszar committed
191
192
193
194
195
196
197
198
199
    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;
200
      t0 = (vp9_dct_value_tokens_ptr + x)->token;
John Koleszar's avatar
John Koleszar committed
201
      /* Consider both possible successor states. */
202
      if (next < default_eob) {
203
        band = band_translate[i + 1];
204
        pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
John Koleszar's avatar
John Koleszar committed
205
        rate0 +=
206
          mb->token_costs[tx_size][type][ref][band][0][pt]
207
                         [tokens[next][0].token];
John Koleszar's avatar
John Koleszar committed
208
        rate1 +=
209
          mb->token_costs[tx_size][type][ref][band][0][pt]
210
                         [tokens[next][1].token];
John Koleszar's avatar
John Koleszar committed
211
      }
212
      UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
213
214
      /* And pick the best. */
      best = rd_cost1 < rd_cost0;
215
      base_bits = *(vp9_dct_value_cost_ptr + x);
216
      dx = mul * (dqcoeff_ptr[rc] - coeff_ptr[rc]);
John Koleszar's avatar
John Koleszar committed
217
218
219
220
221
222
      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;
223
      best_index[i][0] = best;
224

John Koleszar's avatar
John Koleszar committed
225
226
227
228
      /* Evaluate the second possibility for this state. */
      rate0 = tokens[next][0].rate;
      rate1 = tokens[next][1].rate;

229
230
231
      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
232
233
234
235
236
237
238
239
240
241
242
243
244
        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.
245
         */
John Koleszar's avatar
John Koleszar committed
246
247
248
249
250
        t0 = tokens[next][0].token == DCT_EOB_TOKEN ?
             DCT_EOB_TOKEN : ZERO_TOKEN;
        t1 = tokens[next][1].token == DCT_EOB_TOKEN ?
             DCT_EOB_TOKEN : ZERO_TOKEN;
      } else {
251
        t0 = t1 = (vp9_dct_value_tokens_ptr + x)->token;
John Koleszar's avatar
John Koleszar committed
252
      }
253
      if (next < default_eob) {
254
        band = band_translate[i + 1];
John Koleszar's avatar
John Koleszar committed
255
        if (t0 != DCT_EOB_TOKEN) {
256
          pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
257
          rate0 += mb->token_costs[tx_size][type][ref][band][!x][pt]
258
                                  [tokens[next][0].token];
John Koleszar's avatar
John Koleszar committed
259
        }
John Koleszar's avatar
John Koleszar committed
260
        if (t1 != DCT_EOB_TOKEN) {
261
          pt = trellis_get_coeff_context(scan, nb, i, t1, token_cache);
262
          rate1 += mb->token_costs[tx_size][type][ref][band][!x][pt]
263
                                  [tokens[next][1].token];
John Koleszar's avatar
John Koleszar committed
264
265
        }
      }
John Koleszar's avatar
John Koleszar committed
266

267
      UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
268
269
      /* And pick the best. */
      best = rd_cost1 < rd_cost0;
270
      base_bits = *(vp9_dct_value_cost_ptr + x);
John Koleszar's avatar
John Koleszar committed
271
272

      if (shortcut) {
273
        dx -= (dequant_ptr[rc != 0] + sz) ^ sz;
John Koleszar's avatar
John Koleszar committed
274
275
276
277
278
279
280
        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;
281
      best_index[i][1] = best;
John Koleszar's avatar
John Koleszar committed
282
283
      /* Finally, make this the new head of the trellis. */
      next = i;
284
285
286
287
    } 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.
       */
288
      band = band_translate[i + 1];
John Koleszar's avatar
John Koleszar committed
289
290
291
292
      t0 = tokens[next][0].token;
      t1 = tokens[next][1].token;
      /* Update the cost of each path if we're past the EOB token. */
      if (t0 != DCT_EOB_TOKEN) {
293
        tokens[next][0].rate +=
294
            mb->token_costs[tx_size][type][ref][band][1][0][t0];
John Koleszar's avatar
John Koleszar committed
295
296
297
        tokens[next][0].token = ZERO_TOKEN;
      }
      if (t1 != DCT_EOB_TOKEN) {
298
        tokens[next][1].rate +=
299
            mb->token_costs[tx_size][type][ref][band][1][0][t1];
John Koleszar's avatar
John Koleszar committed
300
301
        tokens[next][1].token = ZERO_TOKEN;
      }
302
      best_index[i][0] = best_index[i][1] = 0;
John Koleszar's avatar
John Koleszar committed
303
      /* Don't update next, because we didn't add a new node. */
304
    }
John Koleszar's avatar
John Koleszar committed
305
306
307
  }

  /* Now pick the best path through the whole trellis. */
308
  band = band_translate[i + 1];
309
  pt = combine_entropy_contexts(*a, *l);
John Koleszar's avatar
John Koleszar committed
310
311
312
313
314
315
  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;
316
317
  rate0 += mb->token_costs[tx_size][type][ref][band][0][pt][t0];
  rate1 += mb->token_costs[tx_size][type][ref][band][0][pt][t1];
318
  UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
319
320
  best = rd_cost1 < rd_cost0;
  final_eob = i0 - 1;
321
322
  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
323
324
  for (i = next; i < eob; i = next) {
    x = tokens[i][best].qc;
325
    if (x) {
John Koleszar's avatar
John Koleszar committed
326
      final_eob = i;
327
    }
328
    rc = scan[i];
John Koleszar's avatar
John Koleszar committed
329
    qcoeff_ptr[rc] = x;
330
    dqcoeff_ptr[rc] = (x * dequant_ptr[rc != 0]) / mul;
331

John Koleszar's avatar
John Koleszar committed
332
    next = tokens[i][best].next;
333
    best = best_index[i][best];
John Koleszar's avatar
John Koleszar committed
334
335
336
  }
  final_eob++;

John Koleszar's avatar
John Koleszar committed
337
  xd->plane[plane].eobs[block] = final_eob;
338
  *a = *l = (final_eob > 0);
339
340
}

341
void vp9_optimize_b(int plane, int block, BLOCK_SIZE plane_bsize,
342
                    TX_SIZE tx_size, MACROBLOCK *mb, struct optimize_ctx *ctx) {
John Koleszar's avatar
John Koleszar committed
343
  int x, y;
344
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x, &y);
345
346
  optimize_b(mb, plane, block, plane_bsize,
             &ctx->ta[plane][x], &ctx->tl[plane][y], tx_size);
347
348
}

349
static void optimize_init_b(int plane, BLOCK_SIZE bsize,
350
                            struct encode_b_args *args) {
351
352
  const MACROBLOCKD *xd = &args->x->e_mbd;
  const struct macroblockd_plane* const pd = &xd->plane[plane];
353
  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
354
355
  const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
  const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
356
  const MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
357
  const TX_SIZE tx_size = plane ? get_uv_tx_size(mbmi) : mbmi->tx_size;
358

359
360
361
  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);
362
363
}

364
void vp9_xform_quant(int plane, int block, BLOCK_SIZE plane_bsize,
365
                     TX_SIZE tx_size, void *arg) {
366
367
368
  struct encode_b_args* const args = arg;
  MACROBLOCK* const x = args->x;
  MACROBLOCKD* const xd = &x->e_mbd;
369
370
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
371
  int16_t *coeff = BLOCK_OFFSET(p->coeff, block);
372
  int16_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
373
  int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
374
  const scan_order *so;
375
  uint16_t *eob = &pd->eobs[block];
376
377
  const int diff_stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
  int i, j;
378
  int16_t *src_diff;
379
380
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
  src_diff = &p->src_diff[4 * (j * diff_stride + i)];
381
382

  switch (tx_size) {
383
    case TX_32X32:
384
      so = &vp9_default_scan_orders[TX_32X32];
385
      if (x->use_lp32x32fdct)
386
        vp9_fdct32x32_rd(src_diff, coeff, diff_stride);
387
      else
388
        vp9_fdct32x32(src_diff, coeff, diff_stride);
389
390
      vp9_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round,
                           p->quant, p->quant_shift, qcoeff, dqcoeff,
391
392
                           pd->dequant, p->zbin_extra, eob, so->scan,
                           so->iscan);
393
394
      break;
    case TX_16X16:
395
      so = &vp9_default_scan_orders[TX_16X16];
396
      vp9_fdct16x16(src_diff, coeff, diff_stride);
397
398
      vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
                     p->quant, p->quant_shift, qcoeff, dqcoeff,
399
                     pd->dequant, p->zbin_extra, eob, so->scan, so->iscan);
400
401
      break;
    case TX_8X8:
402
      so = &vp9_default_scan_orders[TX_8X8];
403
      vp9_fdct8x8(src_diff, coeff, diff_stride);
404
405
      vp9_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round,
                     p->quant, p->quant_shift, qcoeff, dqcoeff,
406
                     pd->dequant, p->zbin_extra, eob, so->scan, so->iscan);
407
408
      break;
    case TX_4X4:
409
      so = &vp9_default_scan_orders[TX_4X4];
410
      x->fwd_txm4x4(src_diff, coeff, diff_stride);
411
412
      vp9_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round,
                     p->quant, p->quant_shift, qcoeff, dqcoeff,
413
                     pd->dequant, p->zbin_extra, eob, so->scan, so->iscan);
414
415
416
      break;
    default:
      assert(0);
John Koleszar's avatar
John Koleszar committed
417
  }
418
419
}

420
static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize,
421
                         TX_SIZE tx_size, void *arg) {
422
423
424
  struct encode_b_args *const args = arg;
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
425
  struct optimize_ctx *const ctx = args->ctx;
426
  struct macroblockd_plane *const pd = &xd->plane[plane];
427
  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
428
429
430
431
  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];
432
433
434
435
436

  // 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) {
    pd->eobs[block] = 0;
437
    ctx->ta[plane][i] = 0;
438
    ctx->tl[plane][j] = 0;
439
440
441
    return;
  }

442
  if (!x->skip_recode)
443
    vp9_xform_quant(plane, block, plane_bsize, tx_size, arg);
444

445
  if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
446
    vp9_optimize_b(plane, block, plane_bsize, tx_size, x, ctx);
447
448
  } else {
    ctx->ta[plane][i] = pd->eobs[block] > 0;
449
    ctx->tl[plane][j] = pd->eobs[block] > 0;
450
  }
John Koleszar's avatar
John Koleszar committed
451

452
  if (x->skip_encode || pd->eobs[block] == 0)
453
    return;
454

455
  switch (tx_size) {
456
    case TX_32X32:
Yaowu Xu's avatar
Yaowu Xu committed
457
      vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
458
459
      break;
    case TX_16X16:
460
      vp9_idct16x16_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
461
462
      break;
    case TX_8X8:
463
      vp9_idct8x8_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
464
465
      break;
    case TX_4X4:
466
467
468
      // 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.
469
      xd->itxm_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
470
      break;
471
472
    default:
      assert(!"Invalid transform size");
473
  }
John Koleszar's avatar
John Koleszar committed
474
475
}

476
477
478
479
480
481
482
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;
  struct macroblockd_plane *const pd = &xd->plane[plane];
  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
483
484
485
486
  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];
487
488
489
490
491
492
493
494
495

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

  if (pd->eobs[block] == 0)
    return;

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

496
void vp9_encode_sby(MACROBLOCK *x, BLOCK_SIZE bsize) {
497
  MACROBLOCKD *const xd = &x->e_mbd;
498
  struct optimize_ctx ctx;
499
  struct encode_b_args arg = {x, &ctx};
500
501
502

  vp9_subtract_sby(x, bsize);
  if (x->optimize)
503
    optimize_init_b(0, bsize, &arg);
504

505
  foreach_transformed_block_in_plane(xd, bsize, 0, encode_block_pass1, &arg);
506
507
}

508
void vp9_encode_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
509
  MACROBLOCKD *const xd = &x->e_mbd;
510
  struct optimize_ctx ctx;
511
  struct encode_b_args arg = {x, &ctx};
John Koleszar's avatar
John Koleszar committed
512

513
  if (!x->skip_recode)
514
    vp9_subtract_sb(x, bsize);
515

516
  if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
517
518
519
520
    int i;
    for (i = 0; i < MAX_MB_PLANE; ++i)
      optimize_init_b(i, bsize, &arg);
  }
Scott LaVarnway's avatar
Scott LaVarnway committed
521

Scott LaVarnway's avatar
Scott LaVarnway committed
522
  foreach_transformed_block(xd, bsize, encode_block, &arg);
John Koleszar's avatar
John Koleszar committed
523
}
524

525
void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
526
                            TX_SIZE tx_size, void *arg) {
527
  struct encode_b_args* const args = arg;
528
529
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
530
  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
531
532
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
533
  int16_t *coeff = BLOCK_OFFSET(p->coeff, block);
534
  int16_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
535
  int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
536
  const scan_order *so;
537
  TX_TYPE tx_type;
538
  MB_PREDICTION_MODE mode;
539
540
  const int bwl = b_width_log2(plane_bsize);
  const int diff_stride = 4 * (1 << bwl);
541
542
543
  uint8_t *src, *dst;
  int16_t *src_diff;
  uint16_t *eob = &pd->eobs[block];
544
545
546
547
548
  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)];
549

550
  if (xd->mb_to_right_edge < 0 || xd->mb_to_bottom_edge < 0)
551
    extend_for_intra(xd, plane_bsize, plane, i, j);
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
      so = &vp9_default_scan_orders[TX_32X32];
559
560
561
      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
      block >>= 6;
      vp9_predict_intra_block(xd, block, bwl, TX_32X32, mode,
562
563
564
                              x->skip_encode ? src : dst,
                              x->skip_encode ? p->src.stride : pd->dst.stride,
                              dst, pd->dst.stride);
565
      if (!x->skip_recode) {
566
        vp9_subtract_block(32, 32, src_diff, diff_stride,
567
568
                           src, p->src.stride, dst, pd->dst.stride);
        if (x->use_lp32x32fdct)
569
          vp9_fdct32x32_rd(src_diff, coeff, diff_stride);
570
        else
571
          vp9_fdct32x32(src_diff, coeff, diff_stride);
572
573
        vp9_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round,
                             p->quant, p->quant_shift, qcoeff, dqcoeff,
574
575
                             pd->dequant, p->zbin_extra, eob, so->scan,
                             so->iscan);
576
      }
577
      if (!x->skip_encode && *eob)
Yaowu Xu's avatar
Yaowu Xu committed
578
        vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, *eob);
579
580
      break;
    case TX_16X16:
581
      tx_type = get_tx_type_16x16(pd->plane_type, xd);
582
      so = &vp9_scan_orders[TX_16X16][tx_type];
583
584
585
      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
      block >>= 4;
      vp9_predict_intra_block(xd, block, bwl, TX_16X16, mode,
586
587
588
                              x->skip_encode ? src : dst,
                              x->skip_encode ? p->src.stride : pd->dst.stride,
                              dst, pd->dst.stride);
589
      if (!x->skip_recode) {
590
        vp9_subtract_block(16, 16, src_diff, diff_stride,
591
                           src, p->src.stride, dst, pd->dst.stride);
592
        vp9_fht16x16(tx_type, src_diff, coeff, diff_stride);
593
594
        vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
                       p->quant, p->quant_shift, qcoeff, dqcoeff,
595
                       pd->dequant, p->zbin_extra, eob, so->scan, so->iscan);
596
      }
597
      if (!x->skip_encode && *eob)
598
        vp9_iht16x16_add(tx_type, dqcoeff, dst, pd->dst.stride, *eob);
599
600
      break;
    case TX_8X8:
601
      tx_type = get_tx_type_8x8(pd->plane_type, xd);
602
      so = &vp9_scan_orders[TX_8X8][tx_type];
603
604
605
      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
      block >>= 2;
      vp9_predict_intra_block(xd, block, bwl, TX_8X8, mode,
606
607
608
                              x->skip_encode ? src : dst,
                              x->skip_encode ? p->src.stride : pd->dst.stride,
                              dst, pd->dst.stride);
609
      if (!x->skip_recode) {
610
        vp9_subtract_block(8, 8, src_diff, diff_stride,
611
                           src, p->src.stride, dst, pd->dst.stride);
612
        vp9_fht8x8(tx_type, src_diff, coeff, diff_stride);
613
614
        vp9_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round, p->quant,
                       p->quant_shift, qcoeff, dqcoeff,
615
                       pd->dequant, p->zbin_extra, eob, so->scan, so->iscan);
616
      }
617
      if (!x->skip_encode && *eob)
618
        vp9_iht8x8_add(tx_type, dqcoeff, dst, pd->dst.stride, *eob);
619
620
      break;
    case TX_4X4:
621
      tx_type = get_tx_type_4x4(pd->plane_type, xd, block);
622
      so = &vp9_scan_orders[TX_4X4][tx_type];
623
      if (mbmi->sb_type < BLOCK_8X8 && plane == 0)
624
        mode = xd->mi_8x8[0]->bmi[block].as_mode;
625
      else
626
        mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
627

628
      vp9_predict_intra_block(xd, block, bwl, TX_4X4, mode,
629
630
631
                              x->skip_encode ? src : dst,
                              x->skip_encode ? p->src.stride : pd->dst.stride,
                              dst, pd->dst.stride);
632
633

      if (!x->skip_recode) {
634
        vp9_subtract_block(4, 4, src_diff, diff_stride,
635
636
                           src, p->src.stride, dst, pd->dst.stride);
        if (tx_type != DCT_DCT)
637
          vp9_short_fht4x4(src_diff, coeff, diff_stride, tx_type);
638
        else
639
          x->fwd_txm4x4(src_diff, coeff, diff_stride);
640
641
        vp9_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round, p->quant,
                       p->quant_shift, qcoeff, dqcoeff,
642
                       pd->dequant, p->zbin_extra, eob, so->scan, so->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
658
659
  }
}

660
void vp9_encode_intra_block_y(MACROBLOCK *x, BLOCK_SIZE bsize) {
661
662
  MACROBLOCKD* const xd = &x->e_mbd;
  struct optimize_ctx ctx;
663
  struct encode_b_args arg = {x, &ctx};
664

665
666
  foreach_transformed_block_in_plane(xd, bsize, 0, vp9_encode_block_intra,
                                     &arg);
667
}
668
void vp9_encode_intra_block_uv(MACROBLOCK *x, BLOCK_SIZE bsize) {
669
670
  MACROBLOCKD* const xd = &x->e_mbd;
  struct optimize_ctx ctx;
671
  struct encode_b_args arg = {x, &ctx};
672
  foreach_transformed_block_uv(xd, bsize, vp9_encode_block_intra, &arg);
673
}
674

675
676
677
678
679
680
681
682
683
684
685
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);
}