vp9_encodemb.c 24.6 KB
Newer Older
John Koleszar's avatar
John Koleszar committed
1
/*
2
 *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
John Koleszar's avatar
John Koleszar committed
3
 *
4
 *  Use of this source code is governed by a BSD-style license
5
6
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
7
 *  in the file PATENTS.  All contributing project authors may
8
 *  be found in the AUTHORS file in the root of the source tree.
John Koleszar's avatar
John Koleszar committed
9
10
 */

11
12

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

#include "vpx_mem/vpx_mem.h"

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

22
#include "vp9/encoder/vp9_dct.h"
23
24
25
26
#include "vp9/encoder/vp9_encodemb.h"
#include "vp9/encoder/vp9_quantize.h"
#include "vp9/encoder/vp9_rdopt.h"
#include "vp9/encoder/vp9_tokenize.h"
John Koleszar's avatar
John Koleszar committed
27

28
29
30
31
void vp9_subtract_block_c(int rows, int cols,
                          int16_t *diff_ptr, ptrdiff_t diff_stride,
                          const uint8_t *src_ptr, ptrdiff_t src_stride,
                          const uint8_t *pred_ptr, ptrdiff_t pred_stride) {
John Koleszar's avatar
John Koleszar committed
32
  int r, c;
John Koleszar's avatar
John Koleszar committed
33

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

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

44
static void subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
45
46
47
48
49
50
51
52
53
  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
54
55
}

56
void vp9_subtract_sby(MACROBLOCK *x, BLOCK_SIZE bsize) {
57
  subtract_plane(x, bsize, 0);
58
59
}

60
void vp9_subtract_sbuv(MACROBLOCK *x, BLOCK_SIZE bsize) {
61
  int i;
62

63
64
  for (i = 1; i < MAX_MB_PLANE; i++)
    subtract_plane(x, bsize, i);
65
66
}

67
void vp9_subtract_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
68
69
  vp9_subtract_sby(x, bsize);
  vp9_subtract_sbuv(x, bsize);
John Koleszar's avatar
John Koleszar committed
70
71
}

72
#define RDTRUNC(RM, DM, R, D) ((128 + (R) * (RM)) & 0xFF)
73
typedef struct vp9_token_state vp9_token_state;
74

75
struct vp9_token_state {
76
77
  int           rate;
  int           error;
Daniel Kang's avatar
Daniel Kang committed
78
  int           next;
79
80
81
82
  signed char   token;
  short         qc;
};

83
// TODO(jimbankoski): experiment to find optimal RD numbers.
84
85
#define Y1_RD_MULT 4
#define UV_RD_MULT 2
86

John Koleszar's avatar
John Koleszar committed
87
88
89
static const int plane_rd_mult[4] = {
  Y1_RD_MULT,
  UV_RD_MULT,
90
91
};

92
93
94
95
96
97
98
99
100
101
#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);\
  }\
}

102
103
// This function is a place holder for now but may ultimately need
// to scan previous tokens to work out the correct context.
104
105
static int trellis_get_coeff_context(const int16_t *scan,
                                     const int16_t *nb,
106
                                     int idx, int token,
107
                                     uint8_t *token_cache) {
108
  int bak = token_cache[scan[idx]], pt;
109
  token_cache[scan[idx]] = vp9_pt_energy_class[token];
110
  pt = get_coef_context(nb, token_cache, idx + 1);
111
  token_cache[scan[idx]] = bak;
112
  return pt;
113
114
}

115
static void optimize_b(MACROBLOCK *mb,
116
                       int plane, int block, BLOCK_SIZE plane_bsize,
117
                       ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
John Koleszar's avatar
John Koleszar committed
118
                       TX_SIZE tx_size) {
119
  MACROBLOCKD *const xd = &mb->e_mbd;
120
  struct macroblockd_plane *pd = &xd->plane[plane];
121
  const int ref = is_inter_block(&xd->mi_8x8[0]->mbmi);
122
123
  vp9_token_state tokens[1025][2];
  unsigned best_index[1025][2];
124
  const int16_t *coeff_ptr = BLOCK_OFFSET(mb->plane[plane].coeff, block);
125
126
  int16_t *qcoeff_ptr;
  int16_t *dqcoeff_ptr;
127
  int eob = pd->eobs[block], final_eob, sz = 0;
128
  const int i0 = 0;
129
  int rc, x, next, i;
130
131
132
  int64_t rdmult, rddiv, rd_cost0, rd_cost1;
  int rate0, rate1, error0, error1, t0, t1;
  int best, band, pt;
133
  PLANE_TYPE type = pd->plane_type;
John Koleszar's avatar
John Koleszar committed
134
  int err_mult = plane_rd_mult[type];
135
  const int default_eob = 16 << (tx_size << 1);
136
  const int16_t *scan, *nb;
137
  const int mul = 1 + (tx_size == TX_32X32);
138
  uint8_t token_cache[1024];
139
140
  const int ib = txfrm_block_to_raster_block(plane_bsize, tx_size, block);
  const int16_t *dequant_ptr = pd->dequant;
141
142
143
  const uint8_t *const band_translate = (tx_size == TX_4X4 ?
                                         vp9_coefband_trans_4x4 :
                                         mb->coefband_trans_8x8plus);
John Koleszar's avatar
John Koleszar committed
144

John Koleszar's avatar
John Koleszar committed
145
  assert((!type && !plane) || (type && plane));
146
147
  dqcoeff_ptr = BLOCK_OFFSET(pd->dqcoeff, block);
  qcoeff_ptr = BLOCK_OFFSET(pd->qcoeff, block);
148
  get_scan(xd, tx_size, type, ib, &scan, &nb);
John Koleszar's avatar
John Koleszar committed
149
  assert(eob <= default_eob);
John Koleszar's avatar
John Koleszar committed
150
151
152

  /* Now set up a Viterbi trellis to evaluate alternative roundings. */
  rdmult = mb->rdmult * err_mult;
153
  if (mb->e_mbd.mi_8x8[0]->mbmi.ref_frame[0] == INTRA_FRAME)
John Koleszar's avatar
John Koleszar committed
154
155
156
157
158
    rdmult = (rdmult * 9) >> 4;
  rddiv = mb->rddiv;
  /* Initialize the sentinel node of the trellis. */
  tokens[eob][0].rate = 0;
  tokens[eob][0].error = 0;
159
  tokens[eob][0].next = default_eob;
John Koleszar's avatar
John Koleszar committed
160
161
162
163
  tokens[eob][0].token = DCT_EOB_TOKEN;
  tokens[eob][0].qc = 0;
  *(tokens[eob] + 1) = *(tokens[eob] + 0);
  next = eob;
164
  for (i = 0; i < eob; i++)
165
166
    token_cache[scan[i]] = vp9_pt_energy_class[vp9_dct_value_tokens_ptr[
        qcoeff_ptr[scan[i]]].token];
167

John Koleszar's avatar
John Koleszar committed
168
  for (i = eob; i-- > i0;) {
169
    int base_bits, d2, dx;
John Koleszar's avatar
John Koleszar committed
170

171
    rc = scan[i];
John Koleszar's avatar
John Koleszar committed
172
173
174
175
176
177
178
179
180
    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;
181
      t0 = (vp9_dct_value_tokens_ptr + x)->token;
John Koleszar's avatar
John Koleszar committed
182
      /* Consider both possible successor states. */
183
      if (next < default_eob) {
184
        band = band_translate[i + 1];
185
        pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
John Koleszar's avatar
John Koleszar committed
186
        rate0 +=
187
          mb->token_costs[tx_size][type][ref][band][0][pt]
188
                         [tokens[next][0].token];
John Koleszar's avatar
John Koleszar committed
189
        rate1 +=
190
          mb->token_costs[tx_size][type][ref][band][0][pt]
191
                         [tokens[next][1].token];
John Koleszar's avatar
John Koleszar committed
192
      }
193
      UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
194
195
      /* And pick the best. */
      best = rd_cost1 < rd_cost0;
196
      base_bits = *(vp9_dct_value_cost_ptr + x);
197
      dx = mul * (dqcoeff_ptr[rc] - coeff_ptr[rc]);
John Koleszar's avatar
John Koleszar committed
198
199
200
201
202
203
      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;
204
      best_index[i][0] = best;
205

John Koleszar's avatar
John Koleszar committed
206
207
208
209
      /* Evaluate the second possibility for this state. */
      rate0 = tokens[next][0].rate;
      rate1 = tokens[next][1].rate;

210
211
212
      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
213
214
215
216
217
218
219
220
221
222
223
224
225
        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.
226
         */
John Koleszar's avatar
John Koleszar committed
227
228
229
230
231
        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 {
232
        t0 = t1 = (vp9_dct_value_tokens_ptr + x)->token;
John Koleszar's avatar
John Koleszar committed
233
      }
234
      if (next < default_eob) {
235
        band = band_translate[i + 1];
John Koleszar's avatar
John Koleszar committed
236
        if (t0 != DCT_EOB_TOKEN) {
237
          pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
238
          rate0 += mb->token_costs[tx_size][type][ref][band][!x][pt]
239
                                  [tokens[next][0].token];
John Koleszar's avatar
John Koleszar committed
240
        }
John Koleszar's avatar
John Koleszar committed
241
        if (t1 != DCT_EOB_TOKEN) {
242
          pt = trellis_get_coeff_context(scan, nb, i, t1, token_cache);
243
          rate1 += mb->token_costs[tx_size][type][ref][band][!x][pt]
244
                                  [tokens[next][1].token];
John Koleszar's avatar
John Koleszar committed
245
246
        }
      }
John Koleszar's avatar
John Koleszar committed
247

248
      UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
249
250
      /* And pick the best. */
      best = rd_cost1 < rd_cost0;
251
      base_bits = *(vp9_dct_value_cost_ptr + x);
John Koleszar's avatar
John Koleszar committed
252
253

      if (shortcut) {
254
        dx -= (dequant_ptr[rc != 0] + sz) ^ sz;
John Koleszar's avatar
John Koleszar committed
255
256
257
258
259
260
261
        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;
262
      best_index[i][1] = best;
John Koleszar's avatar
John Koleszar committed
263
264
      /* Finally, make this the new head of the trellis. */
      next = i;
265
266
267
268
    } 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.
       */
269
      band = band_translate[i + 1];
John Koleszar's avatar
John Koleszar committed
270
271
272
273
      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) {
274
        tokens[next][0].rate +=
275
            mb->token_costs[tx_size][type][ref][band][1][0][t0];
John Koleszar's avatar
John Koleszar committed
276
277
278
        tokens[next][0].token = ZERO_TOKEN;
      }
      if (t1 != DCT_EOB_TOKEN) {
279
        tokens[next][1].rate +=
280
            mb->token_costs[tx_size][type][ref][band][1][0][t1];
John Koleszar's avatar
John Koleszar committed
281
282
        tokens[next][1].token = ZERO_TOKEN;
      }
283
      best_index[i][0] = best_index[i][1] = 0;
John Koleszar's avatar
John Koleszar committed
284
      /* Don't update next, because we didn't add a new node. */
285
    }
John Koleszar's avatar
John Koleszar committed
286
287
288
  }

  /* Now pick the best path through the whole trellis. */
289
  band = band_translate[i + 1];
290
  pt = combine_entropy_contexts(*a, *l);
John Koleszar's avatar
John Koleszar committed
291
292
293
294
295
296
  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;
297
298
  rate0 += mb->token_costs[tx_size][type][ref][band][0][pt][t0];
  rate1 += mb->token_costs[tx_size][type][ref][band][0][pt][t1];
299
  UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
300
301
  best = rd_cost1 < rd_cost0;
  final_eob = i0 - 1;
302
303
  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
304
305
  for (i = next; i < eob; i = next) {
    x = tokens[i][best].qc;
306
    if (x) {
John Koleszar's avatar
John Koleszar committed
307
      final_eob = i;
308
    }
309
    rc = scan[i];
John Koleszar's avatar
John Koleszar committed
310
    qcoeff_ptr[rc] = x;
311
    dqcoeff_ptr[rc] = (x * dequant_ptr[rc != 0]) / mul;
312

John Koleszar's avatar
John Koleszar committed
313
    next = tokens[i][best].next;
314
    best = best_index[i][best];
John Koleszar's avatar
John Koleszar committed
315
316
317
  }
  final_eob++;

John Koleszar's avatar
John Koleszar committed
318
  xd->plane[plane].eobs[block] = final_eob;
319
  *a = *l = (final_eob > 0);
320
321
}

322
void vp9_optimize_b(int plane, int block, BLOCK_SIZE plane_bsize,
323
                    TX_SIZE tx_size, MACROBLOCK *mb, struct optimize_ctx *ctx) {
John Koleszar's avatar
John Koleszar committed
324
  int x, y;
325
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x, &y);
326
327
  optimize_b(mb, plane, block, plane_bsize,
             &ctx->ta[plane][x], &ctx->tl[plane][y], tx_size);
328
329
}

330
static void optimize_init_b(int plane, BLOCK_SIZE bsize,
331
                            struct encode_b_args *args) {
332
333
  const MACROBLOCKD *xd = &args->x->e_mbd;
  const struct macroblockd_plane* const pd = &xd->plane[plane];
334
  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
335
336
  const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
  const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
337
  const MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
338
  const TX_SIZE tx_size = plane ? get_uv_tx_size(mbmi) : mbmi->tx_size;
339

340
341
342
  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);
343
344
}

345
void vp9_xform_quant(int plane, int block, BLOCK_SIZE plane_bsize,
346
                     TX_SIZE tx_size, void *arg) {
347
348
349
  struct encode_b_args* const args = arg;
  MACROBLOCK* const x = args->x;
  MACROBLOCKD* const xd = &x->e_mbd;
350
351
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
352
353
354
  int16_t *coeff = BLOCK_OFFSET(p->coeff, block);
  int16_t *qcoeff = BLOCK_OFFSET(pd->qcoeff, block);
  int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
355
356
  const int16_t *scan, *iscan;
  uint16_t *eob = &pd->eobs[block];
357
  const int bwl = b_width_log2(plane_bsize), bw = 1 << bwl;
358
359
360
361
362
  const int twl = bwl - tx_size, twmask = (1 << twl) - 1;
  int xoff, yoff;
  int16_t *src_diff;

  switch (tx_size) {
363
    case TX_32X32:
364
365
366
367
368
369
      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;
370
      if (x->use_lp32x32fdct)
371
        vp9_fdct32x32_rd(src_diff, coeff, bw * 4);
372
      else
373
        vp9_fdct32x32(src_diff, coeff, bw * 4);
374
375
376
      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);
377
378
      break;
    case TX_16X16:
379
380
      scan = vp9_default_scan_16x16;
      iscan = vp9_default_iscan_16x16;
381
382
383
384
      block >>= 4;
      xoff = 16 * (block & twmask);
      yoff = 16 * (block >> twl);
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
385
      vp9_fdct16x16(src_diff, coeff, bw * 4);
386
387
388
      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);
389
390
      break;
    case TX_8X8:
391
392
      scan = vp9_default_scan_8x8;
      iscan = vp9_default_iscan_8x8;
393
394
395
396
      block >>= 2;
      xoff = 8 * (block & twmask);
      yoff = 8 * (block >> twl);
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
397
      vp9_fdct8x8(src_diff, coeff, bw * 4);
398
399
400
      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);
401
402
      break;
    case TX_4X4:
403
404
      scan = vp9_default_scan_4x4;
      iscan = vp9_default_iscan_4x4;
405
406
407
      xoff = 4 * (block & twmask);
      yoff = 4 * (block >> twl);
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
408
      x->fwd_txm4x4(src_diff, coeff, bw * 4);
409
410
411
      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);
412
413
414
      break;
    default:
      assert(0);
John Koleszar's avatar
John Koleszar committed
415
  }
416
417
}

418
static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize,
419
                         TX_SIZE tx_size, void *arg) {
420
421
422
  struct encode_b_args *const args = arg;
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
423
  struct optimize_ctx *const ctx = args->ctx;
424
  struct macroblockd_plane *const pd = &xd->plane[plane];
425
426
427
  const int raster_block = txfrm_block_to_raster_block(plane_bsize, tx_size,
                                                       block);

428
  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
429
  uint8_t *const dst = raster_block_offset_uint8(plane_bsize, raster_block,
430
                                                 pd->dst.buf, pd->dst.stride);
431
432
433
434

  // 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) {
435
    int i, j;
436
    pd->eobs[block] = 0;
437
    txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
438
    ctx->ta[plane][i] = 0;
439
    ctx->tl[plane][j] = 0;
440
441
442
    return;
  }

443
  if (!x->skip_recode)
444
    vp9_xform_quant(plane, block, plane_bsize, tx_size, arg);
445

446
  if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
447
    vp9_optimize_b(plane, block, plane_bsize, tx_size, x, ctx);
448
449
450
451
452
453
  } else {
    int i, k;
    txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &k);
    ctx->ta[plane][i] = pd->eobs[block] > 0;
    ctx->tl[plane][k] = pd->eobs[block] > 0;
  }
John Koleszar's avatar
John Koleszar committed
454

455
  if (x->skip_encode || pd->eobs[block] == 0)
456
    return;
457

458
  switch (tx_size) {
459
    case TX_32X32:
Yaowu Xu's avatar
Yaowu Xu committed
460
      vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
461
462
      break;
    case TX_16X16:
463
      vp9_idct16x16_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
464
465
      break;
    case TX_8X8:
466
      vp9_idct8x8_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
467
468
      break;
    case TX_4X4:
469
470
471
      // 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.
472
      xd->itxm_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
473
      break;
474
475
    default:
      assert(!"Invalid transform size");
476
  }
John Koleszar's avatar
John Koleszar committed
477
478
}

479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
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];
  const int raster_block = txfrm_block_to_raster_block(plane_bsize, tx_size,
                                                       block);

  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
  uint8_t *const dst = raster_block_offset_uint8(plane_bsize, raster_block,
                                                 pd->dst.buf, pd->dst.stride);

  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]);
}

500
void vp9_encode_sby(MACROBLOCK *x, BLOCK_SIZE bsize) {
501
  MACROBLOCKD *const xd = &x->e_mbd;
502
  struct optimize_ctx ctx;
503
  struct encode_b_args arg = {x, &ctx};
504
505
506

  vp9_subtract_sby(x, bsize);
  if (x->optimize)
507
    optimize_init_b(0, bsize, &arg);
508

509
  foreach_transformed_block_in_plane(xd, bsize, 0, encode_block_pass1, &arg);
510
511
}

512
void vp9_encode_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
513
  MACROBLOCKD *const xd = &x->e_mbd;
514
  struct optimize_ctx ctx;
515
  struct encode_b_args arg = {x, &ctx};
John Koleszar's avatar
John Koleszar committed
516

517
  if (!x->skip_recode)
518
    vp9_subtract_sb(x, bsize);
519

520
  if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
521
522
523
524
    int i;
    for (i = 0; i < MAX_MB_PLANE; ++i)
      optimize_init_b(i, bsize, &arg);
  }
Scott LaVarnway's avatar
Scott LaVarnway committed
525

Scott LaVarnway's avatar
Scott LaVarnway committed
526
  foreach_transformed_block(xd, bsize, encode_block, &arg);
John Koleszar's avatar
John Koleszar committed
527
}
528

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

550
  if (xd->mb_to_right_edge < 0 || xd->mb_to_bottom_edge < 0)
551
    extend_for_intra(xd, plane_bsize, plane, block, tx_size);
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
559
560
561
562
563
564
565
566
      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;
      vp9_predict_intra_block(xd, block, bwl, TX_32X32, mode,
                              dst, pd->dst.stride, dst, pd->dst.stride);
567
568
569
570
571
572
573
574
575
576
577
578
579
580

      if (!x->skip_recode) {
        src = p->src.buf + yoff * p->src.stride + xoff;
        src_diff = p->src_diff + 4 * bw * yoff + xoff;
        vp9_subtract_block(32, 32, src_diff, bw * 4,
                           src, p->src.stride, dst, pd->dst.stride);
        if (x->use_lp32x32fdct)
          vp9_fdct32x32_rd(src_diff, coeff, bw * 4);
        else
          vp9_fdct32x32(src_diff, coeff, bw * 4);
        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);
      }
581
      if (!x->skip_encode && *eob)
Yaowu Xu's avatar
Yaowu Xu committed
582
        vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, *eob);
583
584
      break;
    case TX_16X16:
585
      tx_type = get_tx_type_16x16(pd->plane_type, xd);
586
587
588
589
590
591
592
593
594
      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;
      vp9_predict_intra_block(xd, block, bwl, TX_16X16, mode,
                              dst, pd->dst.stride, dst, pd->dst.stride);
595
596
597
598
599
600
601
602
603
604
      if (!x->skip_recode) {
        src = p->src.buf + yoff * p->src.stride + xoff;
        src_diff = p->src_diff + 4 * bw * yoff + xoff;
        vp9_subtract_block(16, 16, src_diff, bw * 4,
                           src, p->src.stride, dst, pd->dst.stride);
        vp9_fht16x16(tx_type, src_diff, coeff, bw * 4);
        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);
      }
605
      if (!x->skip_encode && *eob)
606
        vp9_iht16x16_add(tx_type, dqcoeff, dst, pd->dst.stride, *eob);
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
      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;
      vp9_predict_intra_block(xd, block, bwl, TX_8X8, mode,
                              dst, pd->dst.stride, dst, pd->dst.stride);
619
620
621
622
623
624
625
626
627
628
      if (!x->skip_recode) {
        src = p->src.buf + yoff * p->src.stride + xoff;
        src_diff = p->src_diff + 4 * bw * yoff + xoff;
        vp9_subtract_block(8, 8, src_diff, bw * 4,
                           src, p->src.stride, dst, pd->dst.stride);
        vp9_fht8x8(tx_type, src_diff, coeff, bw * 4);
        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);
      }
629
      if (!x->skip_encode && *eob)
630
        vp9_iht8x8_add(tx_type, dqcoeff, dst, pd->dst.stride, *eob);
631
632
      break;
    case TX_4X4:
633
      tx_type = get_tx_type_4x4(pd->plane_type, xd, block);
634
635
      scan = get_scan_4x4(tx_type);
      iscan = get_iscan_4x4(tx_type);
636
      if (mbmi->sb_type < BLOCK_8X8 && plane == 0)
637
        mode = xd->mi_8x8[0]->bmi[block].as_mode;
638
      else
639
        mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
640

641
642
643
644
645
      xoff = 4 * (block & twmask);
      yoff = 4 * (block >> twl);
      dst = pd->dst.buf + yoff * pd->dst.stride + xoff;
      vp9_predict_intra_block(xd, block, bwl, TX_4X4, mode,
                              dst, pd->dst.stride, dst, pd->dst.stride);
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660

      if (!x->skip_recode) {
        src = p->src.buf + yoff * p->src.stride + xoff;
        src_diff = p->src_diff + 4 * bw * yoff + xoff;
        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);
        else
          x->fwd_txm4x4(src_diff, coeff, bw * 4);
        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);
      }

661
      if (!x->skip_encode && *eob) {
662
663
664
665
        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.
666
          xd->itxm_add(dqcoeff, dst, pd->dst.stride, *eob);
667
        else
668
          vp9_iht4x4_16_add(dqcoeff, dst, pd->dst.stride, tx_type);
669
      }
670
      break;
671
672
    default:
      assert(0);
673
674
675
  }
}

676
void vp9_encode_intra_block_y(MACROBLOCK *x, BLOCK_SIZE bsize) {
677
678
  MACROBLOCKD* const xd = &x->e_mbd;
  struct optimize_ctx ctx;
679
  struct encode_b_args arg = {x, &ctx};
680

681
682
  foreach_transformed_block_in_plane(xd, bsize, 0, vp9_encode_block_intra,
                                     &arg);
683
}
684
void vp9_encode_intra_block_uv(MACROBLOCK *x, BLOCK_SIZE bsize) {
685
686
  MACROBLOCKD* const xd = &x->e_mbd;
  struct optimize_ctx ctx;
687
  struct encode_b_args arg = {x, &ctx};
688
  foreach_transformed_block_uv(xd, bsize, vp9_encode_block_intra, &arg);
689
}
690