vp9_encodemb.c 24.8 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
#include "./vpx_config.h"
12
#include "vp9/encoder/vp9_encodemb.h"
13
#include "vp9/common/vp9_reconinter.h"
14
15
#include "vp9/encoder/vp9_quantize.h"
#include "vp9/encoder/vp9_tokenize.h"
16
#include "vp9/common/vp9_reconintra.h"
John Koleszar's avatar
John Koleszar committed
17
#include "vpx_mem/vpx_mem.h"
18
#include "vp9/encoder/vp9_rdopt.h"
19
#include "vp9/common/vp9_systemdependent.h"
20
#include "vp9_rtcd.h"
John Koleszar's avatar
John Koleszar committed
21

22
23
24
DECLARE_ALIGNED(16, extern const uint8_t,
                vp9_pt_energy_class[MAX_ENTROPY_TOKENS]);

25
26
27
28
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
29
  int r, c;
John Koleszar's avatar
John Koleszar committed
30

31
32
  for (r = 0; r < rows; r++) {
    for (c = 0; c < cols; c++)
John Koleszar's avatar
John Koleszar committed
33
34
      diff_ptr[c] = src_ptr[c] - pred_ptr[c];

35
36
    diff_ptr += diff_stride;
    pred_ptr += pred_stride;
John Koleszar's avatar
John Koleszar committed
37
38
    src_ptr  += src_stride;
  }
John Koleszar's avatar
John Koleszar committed
39
40
}

41
42
43
44
45
46
47
48
49
static void inverse_transform_b_4x4_add(MACROBLOCKD *xd, int eob,
                                        int16_t *dqcoeff, uint8_t *dest,
                                        int stride) {
  if (eob <= 1)
    xd->inv_txm4x4_1_add(dqcoeff, dest, stride);
  else
    xd->inv_txm4x4_add(dqcoeff, dest, stride);
}

50
static void inverse_transform_b_8x8_add(int eob,
51
52
53
54
                                        int16_t *dqcoeff, uint8_t *dest,
                                        int stride) {
  if (eob <= 1)
    vp9_short_idct8x8_1_add(dqcoeff, dest, stride);
55
  else if (eob <= 10)
56
    vp9_short_idct8x8_10_add(dqcoeff, dest, stride);
57
58
59
  else
    vp9_short_idct8x8_add(dqcoeff, dest, stride);
}
60

61
static void inverse_transform_b_16x16_add(int eob,
62
63
                                          int16_t *dqcoeff, uint8_t *dest,
                                          int stride) {
64
65
66
  if (eob <= 1)
    vp9_short_idct16x16_1_add(dqcoeff, dest, stride);
  else if (eob <= 10)
67
    vp9_short_idct16x16_10_add(dqcoeff, dest, stride);
68
69
70
71
  else
    vp9_short_idct16x16_add(dqcoeff, dest, stride);
}

72
static void subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
73
74
75
76
77
78
79
80
81
  struct macroblock_plane *const p = &x->plane[plane];
  const MACROBLOCKD *const xd = &x->e_mbd;
  const struct macroblockd_plane *const pd = &xd->plane[plane];
  const int bw = plane_block_width(bsize, pd);
  const int bh = plane_block_height(bsize, pd);

  vp9_subtract_block(bh, bw, p->src_diff, bw,
                     p->src.buf, p->src.stride,
                     pd->dst.buf, pd->dst.stride);
Yaowu Xu's avatar
Yaowu Xu committed
82
83
}

84
void vp9_subtract_sby(MACROBLOCK *x, BLOCK_SIZE bsize) {
85
  subtract_plane(x, bsize, 0);
86
87
}

88
void vp9_subtract_sbuv(MACROBLOCK *x, BLOCK_SIZE bsize) {
89
  int i;
90

91
92
  for (i = 1; i < MAX_MB_PLANE; i++)
    subtract_plane(x, bsize, i);
93
94
}

95
void vp9_subtract_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
96
97
  vp9_subtract_sby(x, bsize);
  vp9_subtract_sbuv(x, bsize);
John Koleszar's avatar
John Koleszar committed
98
99
}

100

101
#define RDTRUNC(RM,DM,R,D) ( (128+(R)*(RM)) & 0xFF )
102
typedef struct vp9_token_state vp9_token_state;
103

104
struct vp9_token_state {
105
106
  int           rate;
  int           error;
Daniel Kang's avatar
Daniel Kang committed
107
  int           next;
108
109
110
111
  signed char   token;
  short         qc;
};

112
// TODO: experiments to find optimal multiple numbers
113
114
#define Y1_RD_MULT 4
#define UV_RD_MULT 2
115

John Koleszar's avatar
John Koleszar committed
116
117
118
static const int plane_rd_mult[4] = {
  Y1_RD_MULT,
  UV_RD_MULT,
119
120
};

121
122
123
124
125
126
127
128
129
130
#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);\
  }\
}

131
132
// This function is a place holder for now but may ultimately need
// to scan previous tokens to work out the correct context.
133
134
static int trellis_get_coeff_context(const int16_t *scan,
                                     const int16_t *nb,
135
                                     int idx, int token,
136
                                     uint8_t *token_cache) {
137
  int bak = token_cache[scan[idx]], pt;
138
  token_cache[scan[idx]] = vp9_pt_energy_class[token];
139
  pt = get_coef_context(nb, token_cache, idx + 1);
140
  token_cache[scan[idx]] = bak;
141
  return pt;
142
143
}

144
static void optimize_b(MACROBLOCK *mb,
145
                       int plane, int block, BLOCK_SIZE plane_bsize,
146
                       ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
John Koleszar's avatar
John Koleszar committed
147
                       TX_SIZE tx_size) {
148
  MACROBLOCKD *const xd = &mb->e_mbd;
149
  struct macroblockd_plane *pd = &xd->plane[plane];
150
  const int ref = is_inter_block(&xd->this_mi->mbmi);
151
152
  vp9_token_state tokens[1025][2];
  unsigned best_index[1025][2];
153
  const int16_t *coeff_ptr = BLOCK_OFFSET(mb->plane[plane].coeff, block);
154
155
  int16_t *qcoeff_ptr;
  int16_t *dqcoeff_ptr;
156
  int eob = pd->eobs[block], final_eob, sz = 0;
157
  const int i0 = 0;
158
  int rc, x, next, i;
159
160
161
  int64_t rdmult, rddiv, rd_cost0, rd_cost1;
  int rate0, rate1, error0, error1, t0, t1;
  int best, band, pt;
162
  PLANE_TYPE type = pd->plane_type;
John Koleszar's avatar
John Koleszar committed
163
  int err_mult = plane_rd_mult[type];
164
  const int default_eob = 16 << (tx_size << 1);
165
  const int16_t *scan, *nb;
166
  const int mul = 1 + (tx_size == TX_32X32);
167
  uint8_t token_cache[1024];
168
169
  const int ib = txfrm_block_to_raster_block(plane_bsize, tx_size, block);
  const int16_t *dequant_ptr = pd->dequant;
Paul Wilkins's avatar
Paul Wilkins committed
170
  const uint8_t * band_translate;
John Koleszar's avatar
John Koleszar committed
171

John Koleszar's avatar
John Koleszar committed
172
  assert((!type && !plane) || (type && plane));
173
174
  dqcoeff_ptr = BLOCK_OFFSET(pd->dqcoeff, block);
  qcoeff_ptr = BLOCK_OFFSET(pd->qcoeff, block);
175
  get_scan_and_band(xd, tx_size, type, ib, &scan, &band_translate);
John Koleszar's avatar
John Koleszar committed
176
  assert(eob <= default_eob);
John Koleszar's avatar
John Koleszar committed
177
178
179

  /* Now set up a Viterbi trellis to evaluate alternative roundings. */
  rdmult = mb->rdmult * err_mult;
180
  if (mb->e_mbd.mi_8x8[0]->mbmi.ref_frame[0] == INTRA_FRAME)
John Koleszar's avatar
John Koleszar committed
181
182
183
184
185
    rdmult = (rdmult * 9) >> 4;
  rddiv = mb->rddiv;
  /* Initialize the sentinel node of the trellis. */
  tokens[eob][0].rate = 0;
  tokens[eob][0].error = 0;
186
  tokens[eob][0].next = default_eob;
John Koleszar's avatar
John Koleszar committed
187
188
189
190
  tokens[eob][0].token = DCT_EOB_TOKEN;
  tokens[eob][0].qc = 0;
  *(tokens[eob] + 1) = *(tokens[eob] + 0);
  next = eob;
191
  for (i = 0; i < eob; i++)
192
193
    token_cache[scan[i]] = vp9_pt_energy_class[vp9_dct_value_tokens_ptr[
        qcoeff_ptr[scan[i]]].token];
194
  nb = vp9_get_coef_neighbors_handle(scan);
195

John Koleszar's avatar
John Koleszar committed
196
  for (i = eob; i-- > i0;) {
197
    int base_bits, d2, dx;
John Koleszar's avatar
John Koleszar committed
198

199
    rc = scan[i];
John Koleszar's avatar
John Koleszar committed
200
201
202
203
204
205
206
207
208
    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;
209
      t0 = (vp9_dct_value_tokens_ptr + x)->token;
John Koleszar's avatar
John Koleszar committed
210
      /* Consider both possible successor states. */
211
      if (next < default_eob) {
Paul Wilkins's avatar
Paul Wilkins committed
212
        band = get_coef_band(band_translate, i + 1);
213
        pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
John Koleszar's avatar
John Koleszar committed
214
        rate0 +=
215
          mb->token_costs[tx_size][type][ref][band][0][pt]
216
                         [tokens[next][0].token];
John Koleszar's avatar
John Koleszar committed
217
        rate1 +=
218
          mb->token_costs[tx_size][type][ref][band][0][pt]
219
                         [tokens[next][1].token];
John Koleszar's avatar
John Koleszar committed
220
      }
221
      UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
222
223
      /* And pick the best. */
      best = rd_cost1 < rd_cost0;
224
      base_bits = *(vp9_dct_value_cost_ptr + x);
225
      dx = mul * (dqcoeff_ptr[rc] - coeff_ptr[rc]);
John Koleszar's avatar
John Koleszar committed
226
227
228
229
230
231
      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;
232
      best_index[i][0] = best;
233

John Koleszar's avatar
John Koleszar committed
234
235
236
237
      /* Evaluate the second possibility for this state. */
      rate0 = tokens[next][0].rate;
      rate1 = tokens[next][1].rate;

238
239
240
      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
241
242
243
244
245
246
247
248
249
250
251
252
253
        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.
254
         */
John Koleszar's avatar
John Koleszar committed
255
256
257
258
259
        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 {
260
        t0 = t1 = (vp9_dct_value_tokens_ptr + x)->token;
John Koleszar's avatar
John Koleszar committed
261
      }
262
      if (next < default_eob) {
Paul Wilkins's avatar
Paul Wilkins committed
263
        band = get_coef_band(band_translate, i + 1);
John Koleszar's avatar
John Koleszar committed
264
        if (t0 != DCT_EOB_TOKEN) {
265
          pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
266
          rate0 += mb->token_costs[tx_size][type][ref][band][!x][pt]
267
                                  [tokens[next][0].token];
John Koleszar's avatar
John Koleszar committed
268
        }
John Koleszar's avatar
John Koleszar committed
269
        if (t1 != DCT_EOB_TOKEN) {
270
          pt = trellis_get_coeff_context(scan, nb, i, t1, token_cache);
271
          rate1 += mb->token_costs[tx_size][type][ref][band][!x][pt]
272
                                  [tokens[next][1].token];
John Koleszar's avatar
John Koleszar committed
273
274
        }
      }
John Koleszar's avatar
John Koleszar committed
275

276
      UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
277
278
      /* And pick the best. */
      best = rd_cost1 < rd_cost0;
279
      base_bits = *(vp9_dct_value_cost_ptr + x);
John Koleszar's avatar
John Koleszar committed
280
281

      if (shortcut) {
282
        dx -= (dequant_ptr[rc != 0] + sz) ^ sz;
John Koleszar's avatar
John Koleszar committed
283
284
285
286
287
288
289
        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;
290
      best_index[i][1] = best;
John Koleszar's avatar
John Koleszar committed
291
292
      /* Finally, make this the new head of the trellis. */
      next = i;
293
    }
John Koleszar's avatar
John Koleszar committed
294
295
296
297
    /* 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.
     */
    else {
Paul Wilkins's avatar
Paul Wilkins committed
298
      band = get_coef_band(band_translate, i + 1);
John Koleszar's avatar
John Koleszar committed
299
300
301
302
      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) {
303
        tokens[next][0].rate +=
304
            mb->token_costs[tx_size][type][ref][band][1][0][t0];
John Koleszar's avatar
John Koleszar committed
305
306
307
        tokens[next][0].token = ZERO_TOKEN;
      }
      if (t1 != DCT_EOB_TOKEN) {
308
        tokens[next][1].rate +=
309
            mb->token_costs[tx_size][type][ref][band][1][0][t1];
John Koleszar's avatar
John Koleszar committed
310
311
        tokens[next][1].token = ZERO_TOKEN;
      }
312
      best_index[i][0] = best_index[i][1] = 0;
John Koleszar's avatar
John Koleszar committed
313
      /* Don't update next, because we didn't add a new node. */
314
    }
John Koleszar's avatar
John Koleszar committed
315
316
317
  }

  /* Now pick the best path through the whole trellis. */
Paul Wilkins's avatar
Paul Wilkins committed
318
  band = get_coef_band(band_translate, i + 1);
319
  pt = combine_entropy_contexts(*a, *l);
John Koleszar's avatar
John Koleszar committed
320
321
322
323
324
325
  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;
326
327
  rate0 += mb->token_costs[tx_size][type][ref][band][0][pt][t0];
  rate1 += mb->token_costs[tx_size][type][ref][band][0][pt][t1];
328
  UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
329
330
  best = rd_cost1 < rd_cost0;
  final_eob = i0 - 1;
331
332
  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
333
334
  for (i = next; i < eob; i = next) {
    x = tokens[i][best].qc;
335
    if (x) {
John Koleszar's avatar
John Koleszar committed
336
      final_eob = i;
337
    }
338
    rc = scan[i];
John Koleszar's avatar
John Koleszar committed
339
    qcoeff_ptr[rc] = x;
340
    dqcoeff_ptr[rc] = (x * dequant_ptr[rc != 0]) / mul;
341

John Koleszar's avatar
John Koleszar committed
342
    next = tokens[i][best].next;
343
    best = best_index[i][best];
John Koleszar's avatar
John Koleszar committed
344
345
346
  }
  final_eob++;

John Koleszar's avatar
John Koleszar committed
347
  xd->plane[plane].eobs[block] = final_eob;
348
  *a = *l = (final_eob > 0);
349
350
}

351
void vp9_optimize_b(int plane, int block, BLOCK_SIZE plane_bsize,
352
                    TX_SIZE tx_size, MACROBLOCK *mb, struct optimize_ctx *ctx) {
John Koleszar's avatar
John Koleszar committed
353
  int x, y;
354
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x, &y);
355
356
  optimize_b(mb, plane, block, plane_bsize,
             &ctx->ta[plane][x], &ctx->tl[plane][y], tx_size);
357
358
}

359
static void optimize_init_b(int plane, BLOCK_SIZE bsize,
360
                            struct encode_b_args *args) {
361
362
  const MACROBLOCKD *xd = &args->x->e_mbd;
  const struct macroblockd_plane* const pd = &xd->plane[plane];
363
  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
364
365
  const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
  const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
366
  const MB_MODE_INFO *mbmi = &xd->this_mi->mbmi;
367
  const TX_SIZE tx_size = plane ? get_uv_tx_size(mbmi) : mbmi->tx_size;
368

369
370
371
  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);
372
373
}

374
void vp9_xform_quant(int plane, int block, BLOCK_SIZE plane_bsize,
375
                     TX_SIZE tx_size, void *arg) {
376
377
378
  struct encode_b_args* const args = arg;
  MACROBLOCK* const x = args->x;
  MACROBLOCKD* const xd = &x->e_mbd;
379
380
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
381
382
383
  int16_t *coeff = BLOCK_OFFSET(p->coeff, block);
  int16_t *qcoeff = BLOCK_OFFSET(pd->qcoeff, block);
  int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
384
385
  const int16_t *scan, *iscan;
  uint16_t *eob = &pd->eobs[block];
386
  const int bwl = b_width_log2(plane_bsize), bw = 1 << bwl;
387
388
389
390
391
  const int twl = bwl - tx_size, twmask = (1 << twl) - 1;
  int xoff, yoff;
  int16_t *src_diff;

  switch (tx_size) {
392
    case TX_32X32:
393
394
395
396
397
398
      scan = vp9_default_scan_32x32;
      iscan = vp9_default_iscan_32x32;
      block >>= 6;
      xoff = 32 * (block & twmask);
      yoff = 32 * (block >> twl);
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
399
      if (x->use_lp32x32fdct)
400
        vp9_short_fdct32x32_rd(src_diff, coeff, bw * 8);
401
      else
402
403
404
405
        vp9_short_fdct32x32(src_diff, coeff, bw * 8);
      vp9_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round,
                           p->quant, p->quant_shift, qcoeff, dqcoeff,
                           pd->dequant, p->zbin_extra, eob, scan, iscan);
406
407
      break;
    case TX_16X16:
408
409
      scan = vp9_default_scan_16x16;
      iscan = vp9_default_iscan_16x16;
410
411
412
413
      block >>= 4;
      xoff = 16 * (block & twmask);
      yoff = 16 * (block >> twl);
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
414
      x->fwd_txm16x16(src_diff, coeff, bw * 8);
415
416
417
      vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
                     p->quant, p->quant_shift, qcoeff, dqcoeff,
                     pd->dequant, p->zbin_extra, eob, scan, iscan);
418
419
      break;
    case TX_8X8:
420
421
      scan = vp9_default_scan_8x8;
      iscan = vp9_default_iscan_8x8;
422
423
424
425
      block >>= 2;
      xoff = 8 * (block & twmask);
      yoff = 8 * (block >> twl);
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
426
      x->fwd_txm8x8(src_diff, coeff, bw * 8);
427
428
429
      vp9_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round,
                     p->quant, p->quant_shift, qcoeff, dqcoeff,
                     pd->dequant, p->zbin_extra, eob, scan, iscan);
430
431
      break;
    case TX_4X4:
432
433
      scan = vp9_default_scan_4x4;
      iscan = vp9_default_iscan_4x4;
434
435
436
      xoff = 4 * (block & twmask);
      yoff = 4 * (block >> twl);
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
437
      x->fwd_txm4x4(src_diff, coeff, bw * 8);
438
439
440
      vp9_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round,
                     p->quant, p->quant_shift, qcoeff, dqcoeff,
                     pd->dequant, p->zbin_extra, eob, scan, iscan);
441
442
443
      break;
    default:
      assert(0);
John Koleszar's avatar
John Koleszar committed
444
  }
445
446
}

447
static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize,
448
                         TX_SIZE tx_size, void *arg) {
449
450
451
452
  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];
453
454
455
  const int raster_block = txfrm_block_to_raster_block(plane_bsize, tx_size,
                                                       block);

456
  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
457
  uint8_t *const dst = raster_block_offset_uint8(plane_bsize, raster_block,
458
                                                 pd->dst.buf, pd->dst.stride);
459
460
461
462
463
464
465
466

  // 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;
    return;
  }

467
  vp9_xform_quant(plane, block, plane_bsize, tx_size, arg);
468

469
  if (x->optimize)
470
    vp9_optimize_b(plane, block, plane_bsize, tx_size, x, args->ctx);
John Koleszar's avatar
John Koleszar committed
471

472
  if (x->skip_encode || pd->eobs[block] == 0)
473
    return;
474

475
  switch (tx_size) {
476
    case TX_32X32:
477
      vp9_short_idct32x32_add(dqcoeff, dst, pd->dst.stride);
478
479
      break;
    case TX_16X16:
480
481
      inverse_transform_b_16x16_add(pd->eobs[block], dqcoeff, dst,
                                    pd->dst.stride);
482
483
      break;
    case TX_8X8:
484
485
      inverse_transform_b_8x8_add(pd->eobs[block], dqcoeff, dst,
                                  pd->dst.stride);
486
487
      break;
    case TX_4X4:
488
489
490
491
492
      // 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.
      inverse_transform_b_4x4_add(xd, pd->eobs[block], dqcoeff,
                                  dst, pd->dst.stride);
493
      break;
494
495
    default:
      assert(!"Invalid transform size");
496
  }
John Koleszar's avatar
John Koleszar committed
497
498
}

499
void vp9_encode_sby(MACROBLOCK *x, BLOCK_SIZE bsize) {
500
  MACROBLOCKD *const xd = &x->e_mbd;
501
  struct optimize_ctx ctx;
502
  struct encode_b_args arg = {x, &ctx};
503
504
505

  vp9_subtract_sby(x, bsize);
  if (x->optimize)
506
    optimize_init_b(0, bsize, &arg);
507

Scott LaVarnway's avatar
Scott LaVarnway committed
508
  foreach_transformed_block_in_plane(xd, bsize, 0, encode_block, &arg);
509
510
}

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

516
  vp9_subtract_sb(x, bsize);
517
518
519
520
521
522

  if (x->optimize) {
    int i;
    for (i = 0; i < MAX_MB_PLANE; ++i)
      optimize_init_b(i, bsize, &arg);
  }
Scott LaVarnway's avatar
Scott LaVarnway committed
523

Scott LaVarnway's avatar
Scott LaVarnway committed
524
  foreach_transformed_block(xd, bsize, encode_block, &arg);
John Koleszar's avatar
John Koleszar committed
525
}
526

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

548
  if (xd->mb_to_right_edge < 0 || xd->mb_to_bottom_edge < 0)
549
    extend_for_intra(xd, plane_bsize, plane, block, tx_size);
550

551
  // if (x->optimize)
552
  // vp9_optimize_b(plane, block, plane_bsize, tx_size, x, args->ctx);
553

554
  switch (tx_size) {
555
    case TX_32X32:
556
557
558
559
560
561
562
563
564
565
566
567
568
      scan = vp9_default_scan_32x32;
      iscan = vp9_default_iscan_32x32;
      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
      block >>= 6;
      xoff = 32 * (block & twmask);
      yoff = 32 * (block >> twl);
      dst = pd->dst.buf + yoff * pd->dst.stride + xoff;
      src = p->src.buf + yoff * p->src.stride + xoff;
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
      vp9_predict_intra_block(xd, block, bwl, TX_32X32, mode,
                              dst, pd->dst.stride, dst, pd->dst.stride);
      vp9_subtract_block(32, 32, src_diff, bw * 4,
                         src, p->src.stride, dst, pd->dst.stride);
569
      if (x->use_lp32x32fdct)
570
571
572
573
574
575
        vp9_short_fdct32x32_rd(src_diff, coeff, bw * 8);
      else
        vp9_short_fdct32x32(src_diff, coeff, bw * 8);
      vp9_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round,
                           p->quant, p->quant_shift, qcoeff, dqcoeff,
                           pd->dequant, p->zbin_extra, eob, scan, iscan);
576
      if (!x->skip_encode && *eob)
577
        vp9_short_idct32x32_add(dqcoeff, dst, pd->dst.stride);
578
579
      break;
    case TX_16X16:
580
      tx_type = get_tx_type_16x16(pd->plane_type, xd);
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
      scan = get_scan_16x16(tx_type);
      iscan = get_iscan_16x16(tx_type);
      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
      block >>= 4;
      xoff = 16 * (block & twmask);
      yoff = 16 * (block >> twl);
      dst = pd->dst.buf + yoff * pd->dst.stride + xoff;
      src = p->src.buf + yoff * p->src.stride + xoff;
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
      vp9_predict_intra_block(xd, block, bwl, TX_16X16, mode,
                              dst, pd->dst.stride, dst, pd->dst.stride);
      vp9_subtract_block(16, 16, src_diff, bw * 4,
                         src, p->src.stride, dst, pd->dst.stride);
      if (tx_type != DCT_DCT)
        vp9_short_fht16x16(src_diff, coeff, bw * 4, tx_type);
596
      else
597
        x->fwd_txm16x16(src_diff, coeff, bw * 8);
598
599
      vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
                     p->quant, p->quant_shift, qcoeff, dqcoeff,
600
                     pd->dequant, p->zbin_extra, eob, scan, iscan);
601
      if (!x->skip_encode && *eob) {
602
        if (tx_type == DCT_DCT)
603
          inverse_transform_b_16x16_add(*eob, dqcoeff, dst, pd->dst.stride);
604
605
606
        else
          vp9_short_iht16x16_add(dqcoeff, dst, pd->dst.stride, tx_type);
      }
607
608
      break;
    case TX_8X8:
609
      tx_type = get_tx_type_8x8(pd->plane_type, xd);
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
      scan = get_scan_8x8(tx_type);
      iscan = get_iscan_8x8(tx_type);
      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
      block >>= 2;
      xoff = 8 * (block & twmask);
      yoff = 8 * (block >> twl);
      dst = pd->dst.buf + yoff * pd->dst.stride + xoff;
      src = p->src.buf + yoff * p->src.stride + xoff;
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
      vp9_predict_intra_block(xd, block, bwl, TX_8X8, mode,
                              dst, pd->dst.stride, dst, pd->dst.stride);
      vp9_subtract_block(8, 8, src_diff, bw * 4,
                         src, p->src.stride, dst, pd->dst.stride);
      if (tx_type != DCT_DCT)
        vp9_short_fht8x8(src_diff, coeff, bw * 4, tx_type);
625
      else
626
627
628
629
        x->fwd_txm8x8(src_diff, coeff, bw * 8);
      vp9_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round, p->quant,
                     p->quant_shift, qcoeff, dqcoeff,
                     pd->dequant, p->zbin_extra, eob, scan, iscan);
630
      if (!x->skip_encode && *eob) {
631
        if (tx_type == DCT_DCT)
632
          inverse_transform_b_8x8_add(*eob, dqcoeff, dst, pd->dst.stride);
633
634
635
        else
          vp9_short_iht8x8_add(dqcoeff, dst, pd->dst.stride, tx_type);
      }
636
637
      break;
    case TX_4X4:
638
      tx_type = get_tx_type_4x4(pd->plane_type, xd, block);
639
640
      scan = get_scan_4x4(tx_type);
      iscan = get_iscan_4x4(tx_type);
641
      if (mbmi->sb_type < BLOCK_8X8 && plane == 0)
642
        mode = xd->this_mi->bmi[block].as_mode;
643
      else
644
        mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
645

646
647
648
649
650
651
652
653
654
655
656
      xoff = 4 * (block & twmask);
      yoff = 4 * (block >> twl);
      dst = pd->dst.buf + yoff * pd->dst.stride + xoff;
      src = p->src.buf + yoff * p->src.stride + xoff;
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
      vp9_predict_intra_block(xd, block, bwl, TX_4X4, mode,
                              dst, pd->dst.stride, dst, pd->dst.stride);
      vp9_subtract_block(4, 4, src_diff, bw * 4,
                         src, p->src.stride, dst, pd->dst.stride);
      if (tx_type != DCT_DCT)
        vp9_short_fht4x4(src_diff, coeff, bw * 4, tx_type);
657
      else
658
659
660
661
        x->fwd_txm4x4(src_diff, coeff, bw * 8);
      vp9_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round, p->quant,
                     p->quant_shift, qcoeff, dqcoeff,
                     pd->dequant, p->zbin_extra, eob, scan, iscan);
662
      if (!x->skip_encode && *eob) {
663
664
665
666
        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.
667
          inverse_transform_b_4x4_add(xd, *eob, dqcoeff, dst, pd->dst.stride);
668
669
670
        else
          vp9_short_iht4x4_add(dqcoeff, dst, pd->dst.stride, tx_type);
      }
671
      break;
672
673
    default:
      assert(0);
674
675
676
  }
}

677
void vp9_encode_intra_block_y(MACROBLOCK *x, BLOCK_SIZE bsize) {
678
679
  MACROBLOCKD* const xd = &x->e_mbd;
  struct optimize_ctx ctx;
680
  struct encode_b_args arg = {x, &ctx};
681

682
683
  foreach_transformed_block_in_plane(xd, bsize, 0, vp9_encode_block_intra,
                                     &arg);
684
}
685
void vp9_encode_intra_block_uv(MACROBLOCK *x, BLOCK_SIZE bsize) {
686
687
  MACROBLOCKD* const xd = &x->e_mbd;
  struct optimize_ctx ctx;
688
  struct encode_b_args arg = {x, &ctx};
689
  foreach_transformed_block_uv(xd, bsize, vp9_encode_block_intra, &arg);
690
}
691