vp9_encodemb.c 23.7 KB
Newer Older
John Koleszar's avatar
John Koleszar committed
1
/*
2
 *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
John Koleszar's avatar
John Koleszar committed
3
 *
4
 *  Use of this source code is governed by a BSD-style license
5
6
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
7
 *  in the file PATENTS.  All contributing project authors may
8
 *  be found in the AUTHORS file in the root of the source tree.
John Koleszar's avatar
John Koleszar committed
9
10
 */

11
12

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

#include "vpx_mem/vpx_mem.h"

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

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

27
28
29
DECLARE_ALIGNED(16, extern const uint8_t,
                vp9_pt_energy_class[MAX_ENTROPY_TOKENS]);

30
31
32
33
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
34
  int r, c;
John Koleszar's avatar
John Koleszar committed
35

36
37
  for (r = 0; r < rows; r++) {
    for (c = 0; c < cols; c++)
John Koleszar's avatar
John Koleszar committed
38
39
      diff_ptr[c] = src_ptr[c] - pred_ptr[c];

40
41
    diff_ptr += diff_stride;
    pred_ptr += pred_stride;
John Koleszar's avatar
John Koleszar committed
42
43
    src_ptr  += src_stride;
  }
John Koleszar's avatar
John Koleszar committed
44
45
}

46
47
48
49
50
51
52
53
54
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);
}

55
static void subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
56
57
58
59
60
61
62
63
64
  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
65
66
}

67
void vp9_subtract_sby(MACROBLOCK *x, BLOCK_SIZE bsize) {
68
  subtract_plane(x, bsize, 0);
69
70
}

71
void vp9_subtract_sbuv(MACROBLOCK *x, BLOCK_SIZE bsize) {
72
  int i;
73

74
75
  for (i = 1; i < MAX_MB_PLANE; i++)
    subtract_plane(x, bsize, i);
76
77
}

78
void vp9_subtract_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
79
80
  vp9_subtract_sby(x, bsize);
  vp9_subtract_sbuv(x, bsize);
John Koleszar's avatar
John Koleszar committed
81
82
}

83
#define RDTRUNC(RM, DM, R, D) ((128 + (R) * (RM)) & 0xFF)
84
typedef struct vp9_token_state vp9_token_state;
85

86
struct vp9_token_state {
87
88
  int           rate;
  int           error;
Daniel Kang's avatar
Daniel Kang committed
89
  int           next;
90
91
92
93
  signed char   token;
  short         qc;
};

94
// TODO(jimbankoski): experiment to find optimal RD numbers.
95
96
#define Y1_RD_MULT 4
#define UV_RD_MULT 2
97

John Koleszar's avatar
John Koleszar committed
98
99
100
static const int plane_rd_mult[4] = {
  Y1_RD_MULT,
  UV_RD_MULT,
101
102
};

103
104
105
106
107
108
109
110
111
112
#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);\
  }\
}

113
114
// This function is a place holder for now but may ultimately need
// to scan previous tokens to work out the correct context.
115
116
static int trellis_get_coeff_context(const int16_t *scan,
                                     const int16_t *nb,
117
                                     int idx, int token,
118
                                     uint8_t *token_cache) {
119
  int bak = token_cache[scan[idx]], pt;
120
  token_cache[scan[idx]] = vp9_pt_energy_class[token];
121
  pt = get_coef_context(nb, token_cache, idx + 1);
122
  token_cache[scan[idx]] = bak;
123
  return pt;
124
125
}

126
static void optimize_b(MACROBLOCK *mb,
127
                       int plane, int block, BLOCK_SIZE plane_bsize,
128
                       ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
John Koleszar's avatar
John Koleszar committed
129
                       TX_SIZE tx_size) {
130
  MACROBLOCKD *const xd = &mb->e_mbd;
131
  struct macroblockd_plane *pd = &xd->plane[plane];
132
  const int ref = is_inter_block(&xd->this_mi->mbmi);
133
134
  vp9_token_state tokens[1025][2];
  unsigned best_index[1025][2];
135
  const int16_t *coeff_ptr = BLOCK_OFFSET(mb->plane[plane].coeff, block);
136
137
  int16_t *qcoeff_ptr;
  int16_t *dqcoeff_ptr;
138
  int eob = pd->eobs[block], final_eob, sz = 0;
139
  const int i0 = 0;
140
  int rc, x, next, i;
141
142
143
  int64_t rdmult, rddiv, rd_cost0, rd_cost1;
  int rate0, rate1, error0, error1, t0, t1;
  int best, band, pt;
144
  PLANE_TYPE type = pd->plane_type;
John Koleszar's avatar
John Koleszar committed
145
  int err_mult = plane_rd_mult[type];
146
  const int default_eob = 16 << (tx_size << 1);
147
  const int16_t *scan, *nb;
148
  const int mul = 1 + (tx_size == TX_32X32);
149
  uint8_t token_cache[1024];
150
151
  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
152
  const uint8_t * band_translate;
John Koleszar's avatar
John Koleszar committed
153

John Koleszar's avatar
John Koleszar committed
154
  assert((!type && !plane) || (type && plane));
155
156
  dqcoeff_ptr = BLOCK_OFFSET(pd->dqcoeff, block);
  qcoeff_ptr = BLOCK_OFFSET(pd->qcoeff, block);
157
  get_scan_and_band(xd, tx_size, type, ib, &scan, &nb, &band_translate);
John Koleszar's avatar
John Koleszar committed
158
  assert(eob <= default_eob);
John Koleszar's avatar
John Koleszar committed
159
160
161

  /* Now set up a Viterbi trellis to evaluate alternative roundings. */
  rdmult = mb->rdmult * err_mult;
162
  if (mb->e_mbd.mi_8x8[0]->mbmi.ref_frame[0] == INTRA_FRAME)
John Koleszar's avatar
John Koleszar committed
163
164
165
166
167
    rdmult = (rdmult * 9) >> 4;
  rddiv = mb->rddiv;
  /* Initialize the sentinel node of the trellis. */
  tokens[eob][0].rate = 0;
  tokens[eob][0].error = 0;
168
  tokens[eob][0].next = default_eob;
John Koleszar's avatar
John Koleszar committed
169
170
171
172
  tokens[eob][0].token = DCT_EOB_TOKEN;
  tokens[eob][0].qc = 0;
  *(tokens[eob] + 1) = *(tokens[eob] + 0);
  next = eob;
173
  for (i = 0; i < eob; i++)
174
175
    token_cache[scan[i]] = vp9_pt_energy_class[vp9_dct_value_tokens_ptr[
        qcoeff_ptr[scan[i]]].token];
176

John Koleszar's avatar
John Koleszar committed
177
  for (i = eob; i-- > i0;) {
178
    int base_bits, d2, dx;
John Koleszar's avatar
John Koleszar committed
179

180
    rc = scan[i];
John Koleszar's avatar
John Koleszar committed
181
182
183
184
185
186
187
188
189
    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;
190
      t0 = (vp9_dct_value_tokens_ptr + x)->token;
John Koleszar's avatar
John Koleszar committed
191
      /* Consider both possible successor states. */
192
      if (next < default_eob) {
Paul Wilkins's avatar
Paul Wilkins committed
193
        band = get_coef_band(band_translate, i + 1);
194
        pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
John Koleszar's avatar
John Koleszar committed
195
        rate0 +=
196
          mb->token_costs[tx_size][type][ref][band][0][pt]
197
                         [tokens[next][0].token];
John Koleszar's avatar
John Koleszar committed
198
        rate1 +=
199
          mb->token_costs[tx_size][type][ref][band][0][pt]
200
                         [tokens[next][1].token];
John Koleszar's avatar
John Koleszar committed
201
      }
202
      UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
203
204
      /* And pick the best. */
      best = rd_cost1 < rd_cost0;
205
      base_bits = *(vp9_dct_value_cost_ptr + x);
206
      dx = mul * (dqcoeff_ptr[rc] - coeff_ptr[rc]);
John Koleszar's avatar
John Koleszar committed
207
208
209
210
211
212
      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;
213
      best_index[i][0] = best;
214

John Koleszar's avatar
John Koleszar committed
215
216
217
218
      /* Evaluate the second possibility for this state. */
      rate0 = tokens[next][0].rate;
      rate1 = tokens[next][1].rate;

219
220
221
      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
222
223
224
225
226
227
228
229
230
231
232
233
234
        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.
235
         */
John Koleszar's avatar
John Koleszar committed
236
237
238
239
240
        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 {
241
        t0 = t1 = (vp9_dct_value_tokens_ptr + x)->token;
John Koleszar's avatar
John Koleszar committed
242
      }
243
      if (next < default_eob) {
Paul Wilkins's avatar
Paul Wilkins committed
244
        band = get_coef_band(band_translate, i + 1);
John Koleszar's avatar
John Koleszar committed
245
        if (t0 != DCT_EOB_TOKEN) {
246
          pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
247
          rate0 += mb->token_costs[tx_size][type][ref][band][!x][pt]
248
                                  [tokens[next][0].token];
John Koleszar's avatar
John Koleszar committed
249
        }
John Koleszar's avatar
John Koleszar committed
250
        if (t1 != DCT_EOB_TOKEN) {
251
          pt = trellis_get_coeff_context(scan, nb, i, t1, token_cache);
252
          rate1 += mb->token_costs[tx_size][type][ref][band][!x][pt]
253
                                  [tokens[next][1].token];
John Koleszar's avatar
John Koleszar committed
254
255
        }
      }
John Koleszar's avatar
John Koleszar committed
256

257
      UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
258
259
      /* And pick the best. */
      best = rd_cost1 < rd_cost0;
260
      base_bits = *(vp9_dct_value_cost_ptr + x);
John Koleszar's avatar
John Koleszar committed
261
262

      if (shortcut) {
263
        dx -= (dequant_ptr[rc != 0] + sz) ^ sz;
John Koleszar's avatar
John Koleszar committed
264
265
266
267
268
269
270
        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;
271
      best_index[i][1] = best;
John Koleszar's avatar
John Koleszar committed
272
273
      /* Finally, make this the new head of the trellis. */
      next = i;
274
275
276
277
    } 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.
       */
Paul Wilkins's avatar
Paul Wilkins committed
278
      band = get_coef_band(band_translate, i + 1);
John Koleszar's avatar
John Koleszar committed
279
280
281
282
      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) {
283
        tokens[next][0].rate +=
284
            mb->token_costs[tx_size][type][ref][band][1][0][t0];
John Koleszar's avatar
John Koleszar committed
285
286
287
        tokens[next][0].token = ZERO_TOKEN;
      }
      if (t1 != DCT_EOB_TOKEN) {
288
        tokens[next][1].rate +=
289
            mb->token_costs[tx_size][type][ref][band][1][0][t1];
John Koleszar's avatar
John Koleszar committed
290
291
        tokens[next][1].token = ZERO_TOKEN;
      }
292
      best_index[i][0] = best_index[i][1] = 0;
John Koleszar's avatar
John Koleszar committed
293
      /* Don't update next, because we didn't add a new node. */
294
    }
John Koleszar's avatar
John Koleszar committed
295
296
297
  }

  /* Now pick the best path through the whole trellis. */
Paul Wilkins's avatar
Paul Wilkins committed
298
  band = get_coef_band(band_translate, i + 1);
299
  pt = combine_entropy_contexts(*a, *l);
John Koleszar's avatar
John Koleszar committed
300
301
302
303
304
305
  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;
306
307
  rate0 += mb->token_costs[tx_size][type][ref][band][0][pt][t0];
  rate1 += mb->token_costs[tx_size][type][ref][band][0][pt][t1];
308
  UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
309
310
  best = rd_cost1 < rd_cost0;
  final_eob = i0 - 1;
311
312
  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
313
314
  for (i = next; i < eob; i = next) {
    x = tokens[i][best].qc;
315
    if (x) {
John Koleszar's avatar
John Koleszar committed
316
      final_eob = i;
317
    }
318
    rc = scan[i];
John Koleszar's avatar
John Koleszar committed
319
    qcoeff_ptr[rc] = x;
320
    dqcoeff_ptr[rc] = (x * dequant_ptr[rc != 0]) / mul;
321

John Koleszar's avatar
John Koleszar committed
322
    next = tokens[i][best].next;
323
    best = best_index[i][best];
John Koleszar's avatar
John Koleszar committed
324
325
326
  }
  final_eob++;

John Koleszar's avatar
John Koleszar committed
327
  xd->plane[plane].eobs[block] = final_eob;
328
  *a = *l = (final_eob > 0);
329
330
}

331
void vp9_optimize_b(int plane, int block, BLOCK_SIZE plane_bsize,
332
                    TX_SIZE tx_size, MACROBLOCK *mb, struct optimize_ctx *ctx) {
John Koleszar's avatar
John Koleszar committed
333
  int x, y;
334
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x, &y);
335
336
  optimize_b(mb, plane, block, plane_bsize,
             &ctx->ta[plane][x], &ctx->tl[plane][y], tx_size);
337
338
}

339
static void optimize_init_b(int plane, BLOCK_SIZE bsize,
340
                            struct encode_b_args *args) {
341
342
  const MACROBLOCKD *xd = &args->x->e_mbd;
  const struct macroblockd_plane* const pd = &xd->plane[plane];
343
  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
344
345
  const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
  const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
346
  const MB_MODE_INFO *mbmi = &xd->this_mi->mbmi;
347
  const TX_SIZE tx_size = plane ? get_uv_tx_size(mbmi) : mbmi->tx_size;
348

349
350
351
  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);
352
353
}

354
void vp9_xform_quant(int plane, int block, BLOCK_SIZE plane_bsize,
355
                     TX_SIZE tx_size, void *arg) {
356
357
358
  struct encode_b_args* const args = arg;
  MACROBLOCK* const x = args->x;
  MACROBLOCKD* const xd = &x->e_mbd;
359
360
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
361
362
363
  int16_t *coeff = BLOCK_OFFSET(p->coeff, block);
  int16_t *qcoeff = BLOCK_OFFSET(pd->qcoeff, block);
  int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
364
365
  const int16_t *scan, *iscan;
  uint16_t *eob = &pd->eobs[block];
366
  const int bwl = b_width_log2(plane_bsize), bw = 1 << bwl;
367
368
369
370
371
  const int twl = bwl - tx_size, twmask = (1 << twl) - 1;
  int xoff, yoff;
  int16_t *src_diff;

  switch (tx_size) {
372
    case TX_32X32:
373
374
375
376
377
378
      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;
379
      if (x->use_lp32x32fdct)
380
        vp9_short_fdct32x32_rd(src_diff, coeff, bw * 8);
381
      else
382
383
384
385
        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);
386
387
      break;
    case TX_16X16:
388
389
      scan = vp9_default_scan_16x16;
      iscan = vp9_default_iscan_16x16;
390
391
392
393
      block >>= 4;
      xoff = 16 * (block & twmask);
      yoff = 16 * (block >> twl);
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
394
      x->fwd_txm16x16(src_diff, coeff, bw * 8);
395
396
397
      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);
398
399
      break;
    case TX_8X8:
400
401
      scan = vp9_default_scan_8x8;
      iscan = vp9_default_iscan_8x8;
402
403
404
405
      block >>= 2;
      xoff = 8 * (block & twmask);
      yoff = 8 * (block >> twl);
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
406
      x->fwd_txm8x8(src_diff, coeff, bw * 8);
407
408
409
      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);
410
411
      break;
    case TX_4X4:
412
413
      scan = vp9_default_scan_4x4;
      iscan = vp9_default_iscan_4x4;
414
415
416
      xoff = 4 * (block & twmask);
      yoff = 4 * (block >> twl);
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
417
      x->fwd_txm4x4(src_diff, coeff, bw * 8);
418
419
420
      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);
421
422
423
      break;
    default:
      assert(0);
John Koleszar's avatar
John Koleszar committed
424
  }
425
426
}

427
static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize,
428
                         TX_SIZE tx_size, void *arg) {
429
430
431
432
  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];
433
434
435
  const int raster_block = txfrm_block_to_raster_block(plane_bsize, tx_size,
                                                       block);

436
  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
437
  uint8_t *const dst = raster_block_offset_uint8(plane_bsize, raster_block,
438
                                                 pd->dst.buf, pd->dst.stride);
439
440
441
442
443
444
445
446

  // 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;
  }

447
  vp9_xform_quant(plane, block, plane_bsize, tx_size, arg);
448

449
  if (x->optimize)
450
    vp9_optimize_b(plane, block, plane_bsize, tx_size, x, args->ctx);
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:
457
      vp9_short_idct32x32_add(dqcoeff, dst, pd->dst.stride);
458
459
      break;
    case TX_16X16:
460
      vp9_idct_add_16x16(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
461
462
      break;
    case TX_8X8:
463
      vp9_idct_add_8x8(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
464
465
      break;
    case TX_4X4:
466
467
468
469
470
      // 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);
471
      break;
472
473
    default:
      assert(!"Invalid transform size");
474
  }
John Koleszar's avatar
John Koleszar committed
475
476
}

477
void vp9_encode_sby(MACROBLOCK *x, BLOCK_SIZE bsize) {
478
  MACROBLOCKD *const xd = &x->e_mbd;
479
  struct optimize_ctx ctx;
480
  struct encode_b_args arg = {x, &ctx};
481
482
483

  vp9_subtract_sby(x, bsize);
  if (x->optimize)
484
    optimize_init_b(0, bsize, &arg);
485

Scott LaVarnway's avatar
Scott LaVarnway committed
486
  foreach_transformed_block_in_plane(xd, bsize, 0, encode_block, &arg);
487
488
}

489
void vp9_encode_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
490
  MACROBLOCKD *const xd = &x->e_mbd;
491
  struct optimize_ctx ctx;
492
  struct encode_b_args arg = {x, &ctx};
John Koleszar's avatar
John Koleszar committed
493

494
  vp9_subtract_sb(x, bsize);
495
496
497
498
499
500

  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
501

Scott LaVarnway's avatar
Scott LaVarnway committed
502
  foreach_transformed_block(xd, bsize, encode_block, &arg);
John Koleszar's avatar
John Koleszar committed
503
}
504

505
void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
506
                            TX_SIZE tx_size, void *arg) {
507
  struct encode_b_args* const args = arg;
508
509
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
510
  MB_MODE_INFO *mbmi = &xd->this_mi->mbmi;
511
512
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
513
514
515
  int16_t *coeff = BLOCK_OFFSET(p->coeff, block);
  int16_t *qcoeff = BLOCK_OFFSET(pd->qcoeff, block);
  int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
516
  const int16_t *scan, *iscan;
517
  TX_TYPE tx_type;
518
  MB_PREDICTION_MODE mode;
519
  const int bwl = b_width_log2(plane_bsize), bw = 1 << bwl;
520
521
522
523
524
  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];
525

526
  if (xd->mb_to_right_edge < 0 || xd->mb_to_bottom_edge < 0)
527
    extend_for_intra(xd, plane_bsize, plane, block, tx_size);
528

529
  // if (x->optimize)
530
  // vp9_optimize_b(plane, block, plane_bsize, tx_size, x, args->ctx);
531

532
  switch (tx_size) {
533
    case TX_32X32:
534
535
536
537
538
539
540
541
542
543
544
545
546
      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);
547
      if (x->use_lp32x32fdct)
548
549
550
551
552
553
        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);
554
      if (!x->skip_encode && *eob)
555
        vp9_short_idct32x32_add(dqcoeff, dst, pd->dst.stride);
556
557
      break;
    case TX_16X16:
558
      tx_type = get_tx_type_16x16(pd->plane_type, xd);
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
      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);
574
      else
575
        x->fwd_txm16x16(src_diff, coeff, bw * 8);
576
577
      vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
                     p->quant, p->quant_shift, qcoeff, dqcoeff,
578
                     pd->dequant, p->zbin_extra, eob, scan, iscan);
579
580
      if (!x->skip_encode && *eob)
        vp9_iht_add_16x16(tx_type, dqcoeff, dst, pd->dst.stride, *eob);
581
582
      break;
    case TX_8X8:
583
      tx_type = get_tx_type_8x8(pd->plane_type, xd);
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
      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);
599
      else
600
601
602
603
        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);
604
605
      if (!x->skip_encode && *eob)
        vp9_iht_add_8x8(tx_type, dqcoeff, dst, pd->dst.stride, *eob);
606
607
      break;
    case TX_4X4:
608
      tx_type = get_tx_type_4x4(pd->plane_type, xd, block);
609
610
      scan = get_scan_4x4(tx_type);
      iscan = get_iscan_4x4(tx_type);
611
      if (mbmi->sb_type < BLOCK_8X8 && plane == 0)
612
        mode = xd->this_mi->bmi[block].as_mode;
613
      else
614
        mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
615

616
617
618
619
620
621
622
623
624
625
626
      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);
627
      else
628
629
630
631
        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);
632
      if (!x->skip_encode && *eob) {
633
634
635
636
        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.
637
          inverse_transform_b_4x4_add(xd, *eob, dqcoeff, dst, pd->dst.stride);
638
639
640
        else
          vp9_short_iht4x4_add(dqcoeff, dst, pd->dst.stride, tx_type);
      }
641
      break;
642
643
    default:
      assert(0);
644
645
646
  }
}

647
void vp9_encode_intra_block_y(MACROBLOCK *x, BLOCK_SIZE bsize) {
648
649
  MACROBLOCKD* const xd = &x->e_mbd;
  struct optimize_ctx ctx;
650
  struct encode_b_args arg = {x, &ctx};
651

652
653
  foreach_transformed_block_in_plane(xd, bsize, 0, vp9_encode_block_intra,
                                     &arg);
654
}
655
void vp9_encode_intra_block_uv(MACROBLOCK *x, BLOCK_SIZE bsize) {
656
657
  MACROBLOCKD* const xd = &x->e_mbd;
  struct optimize_ctx ctx;
658
  struct encode_b_args arg = {x, &ctx};
659
  foreach_transformed_block_uv(xd, bsize, vp9_encode_block_intra, &arg);
660
}
661