vp9_encodemb.c 24.3 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
  const int16_t *dequant_ptr = pd->dequant;
140
141
142
  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
143

John Koleszar's avatar
John Koleszar committed
144
  assert((!type && !plane) || (type && plane));
145
146
  dqcoeff_ptr = BLOCK_OFFSET(pd->dqcoeff, block);
  qcoeff_ptr = BLOCK_OFFSET(pd->qcoeff, block);
147
  get_scan(xd, tx_size, type, block, &scan, &nb);
John Koleszar's avatar
John Koleszar committed
148
  assert(eob <= default_eob);
John Koleszar's avatar
John Koleszar committed
149
150
151

  /* Now set up a Viterbi trellis to evaluate alternative roundings. */
  rdmult = mb->rdmult * err_mult;
152
  if (mb->e_mbd.mi_8x8[0]->mbmi.ref_frame[0] == INTRA_FRAME)
John Koleszar's avatar
John Koleszar committed
153
154
155
156
157
    rdmult = (rdmult * 9) >> 4;
  rddiv = mb->rddiv;
  /* Initialize the sentinel node of the trellis. */
  tokens[eob][0].rate = 0;
  tokens[eob][0].error = 0;
158
  tokens[eob][0].next = default_eob;
John Koleszar's avatar
John Koleszar committed
159
160
161
162
  tokens[eob][0].token = DCT_EOB_TOKEN;
  tokens[eob][0].qc = 0;
  *(tokens[eob] + 1) = *(tokens[eob] + 0);
  next = eob;
163
  for (i = 0; i < eob; i++)
164
165
    token_cache[scan[i]] = vp9_pt_energy_class[vp9_dct_value_tokens_ptr[
        qcoeff_ptr[scan[i]]].token];
166

John Koleszar's avatar
John Koleszar committed
167
  for (i = eob; i-- > i0;) {
168
    int base_bits, d2, dx;
John Koleszar's avatar
John Koleszar committed
169

170
    rc = scan[i];
John Koleszar's avatar
John Koleszar committed
171
172
173
174
175
176
177
178
179
    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;
180
      t0 = (vp9_dct_value_tokens_ptr + x)->token;
John Koleszar's avatar
John Koleszar committed
181
      /* Consider both possible successor states. */
182
      if (next < default_eob) {
183
        band = band_translate[i + 1];
184
        pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
John Koleszar's avatar
John Koleszar committed
185
        rate0 +=
186
          mb->token_costs[tx_size][type][ref][band][0][pt]
187
                         [tokens[next][0].token];
John Koleszar's avatar
John Koleszar committed
188
        rate1 +=
189
          mb->token_costs[tx_size][type][ref][band][0][pt]
190
                         [tokens[next][1].token];
John Koleszar's avatar
John Koleszar committed
191
      }
192
      UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
193
194
      /* And pick the best. */
      best = rd_cost1 < rd_cost0;
195
      base_bits = *(vp9_dct_value_cost_ptr + x);
196
      dx = mul * (dqcoeff_ptr[rc] - coeff_ptr[rc]);
John Koleszar's avatar
John Koleszar committed
197
198
199
200
201
202
      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;
203
      best_index[i][0] = best;
204

John Koleszar's avatar
John Koleszar committed
205
206
207
208
      /* Evaluate the second possibility for this state. */
      rate0 = tokens[next][0].rate;
      rate1 = tokens[next][1].rate;

209
210
211
      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
212
213
214
215
216
217
218
219
220
221
222
223
224
        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.
225
         */
John Koleszar's avatar
John Koleszar committed
226
227
228
229
230
        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 {
231
        t0 = t1 = (vp9_dct_value_tokens_ptr + x)->token;
John Koleszar's avatar
John Koleszar committed
232
      }
233
      if (next < default_eob) {
234
        band = band_translate[i + 1];
John Koleszar's avatar
John Koleszar committed
235
        if (t0 != DCT_EOB_TOKEN) {
236
          pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
237
          rate0 += mb->token_costs[tx_size][type][ref][band][!x][pt]
238
                                  [tokens[next][0].token];
John Koleszar's avatar
John Koleszar committed
239
        }
John Koleszar's avatar
John Koleszar committed
240
        if (t1 != DCT_EOB_TOKEN) {
241
          pt = trellis_get_coeff_context(scan, nb, i, t1, token_cache);
242
          rate1 += mb->token_costs[tx_size][type][ref][band][!x][pt]
243
                                  [tokens[next][1].token];
John Koleszar's avatar
John Koleszar committed
244
245
        }
      }
John Koleszar's avatar
John Koleszar committed
246

247
      UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
248
249
      /* And pick the best. */
      best = rd_cost1 < rd_cost0;
250
      base_bits = *(vp9_dct_value_cost_ptr + x);
John Koleszar's avatar
John Koleszar committed
251
252

      if (shortcut) {
253
        dx -= (dequant_ptr[rc != 0] + sz) ^ sz;
John Koleszar's avatar
John Koleszar committed
254
255
256
257
258
259
260
        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;
261
      best_index[i][1] = best;
John Koleszar's avatar
John Koleszar committed
262
263
      /* Finally, make this the new head of the trellis. */
      next = i;
264
265
266
267
    } 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.
       */
268
      band = band_translate[i + 1];
John Koleszar's avatar
John Koleszar committed
269
270
271
272
      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) {
273
        tokens[next][0].rate +=
274
            mb->token_costs[tx_size][type][ref][band][1][0][t0];
John Koleszar's avatar
John Koleszar committed
275
276
277
        tokens[next][0].token = ZERO_TOKEN;
      }
      if (t1 != DCT_EOB_TOKEN) {
278
        tokens[next][1].rate +=
279
            mb->token_costs[tx_size][type][ref][band][1][0][t1];
John Koleszar's avatar
John Koleszar committed
280
281
        tokens[next][1].token = ZERO_TOKEN;
      }
282
      best_index[i][0] = best_index[i][1] = 0;
John Koleszar's avatar
John Koleszar committed
283
      /* Don't update next, because we didn't add a new node. */
284
    }
John Koleszar's avatar
John Koleszar committed
285
286
287
  }

  /* Now pick the best path through the whole trellis. */
288
  band = band_translate[i + 1];
289
  pt = combine_entropy_contexts(*a, *l);
John Koleszar's avatar
John Koleszar committed
290
291
292
293
294
295
  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;
296
297
  rate0 += mb->token_costs[tx_size][type][ref][band][0][pt][t0];
  rate1 += mb->token_costs[tx_size][type][ref][band][0][pt][t1];
298
  UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
299
300
  best = rd_cost1 < rd_cost0;
  final_eob = i0 - 1;
301
302
  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
303
304
  for (i = next; i < eob; i = next) {
    x = tokens[i][best].qc;
305
    if (x) {
John Koleszar's avatar
John Koleszar committed
306
      final_eob = i;
307
    }
308
    rc = scan[i];
John Koleszar's avatar
John Koleszar committed
309
    qcoeff_ptr[rc] = x;
310
    dqcoeff_ptr[rc] = (x * dequant_ptr[rc != 0]) / mul;
311

John Koleszar's avatar
John Koleszar committed
312
    next = tokens[i][best].next;
313
    best = best_index[i][best];
John Koleszar's avatar
John Koleszar committed
314
315
316
  }
  final_eob++;

John Koleszar's avatar
John Koleszar committed
317
  xd->plane[plane].eobs[block] = final_eob;
318
  *a = *l = (final_eob > 0);
319
320
}

321
void vp9_optimize_b(int plane, int block, BLOCK_SIZE plane_bsize,
322
                    TX_SIZE tx_size, MACROBLOCK *mb, struct optimize_ctx *ctx) {
John Koleszar's avatar
John Koleszar committed
323
  int x, y;
324
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x, &y);
325
326
  optimize_b(mb, plane, block, plane_bsize,
             &ctx->ta[plane][x], &ctx->tl[plane][y], tx_size);
327
328
}

329
static void optimize_init_b(int plane, BLOCK_SIZE bsize,
330
                            struct encode_b_args *args) {
331
332
  const MACROBLOCKD *xd = &args->x->e_mbd;
  const struct macroblockd_plane* const pd = &xd->plane[plane];
333
  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
334
335
  const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
  const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
336
  const MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
337
  const TX_SIZE tx_size = plane ? get_uv_tx_size(mbmi) : mbmi->tx_size;
338

339
340
341
  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);
342
343
}

344
void vp9_xform_quant(int plane, int block, BLOCK_SIZE plane_bsize,
345
                     TX_SIZE tx_size, void *arg) {
346
347
348
  struct encode_b_args* const args = arg;
  MACROBLOCK* const x = args->x;
  MACROBLOCKD* const xd = &x->e_mbd;
349
350
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
351
352
353
  int16_t *coeff = BLOCK_OFFSET(p->coeff, block);
  int16_t *qcoeff = BLOCK_OFFSET(pd->qcoeff, block);
  int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
354
355
  const int16_t *scan, *iscan;
  uint16_t *eob = &pd->eobs[block];
356
  const int bwl = b_width_log2(plane_bsize), bw = 1 << bwl;
357
358
359
360
361
  const int twl = bwl - tx_size, twmask = (1 << twl) - 1;
  int xoff, yoff;
  int16_t *src_diff;

  switch (tx_size) {
362
    case TX_32X32:
363
364
365
366
367
368
      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;
369
      if (x->use_lp32x32fdct)
370
        vp9_fdct32x32_rd(src_diff, coeff, bw * 4);
371
      else
372
        vp9_fdct32x32(src_diff, coeff, bw * 4);
373
374
375
      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);
376
377
      break;
    case TX_16X16:
378
379
      scan = vp9_default_scan_16x16;
      iscan = vp9_default_iscan_16x16;
380
381
382
383
      block >>= 4;
      xoff = 16 * (block & twmask);
      yoff = 16 * (block >> twl);
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
384
      vp9_fdct16x16(src_diff, coeff, bw * 4);
385
386
387
      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);
388
389
      break;
    case TX_8X8:
390
391
      scan = vp9_default_scan_8x8;
      iscan = vp9_default_iscan_8x8;
392
393
394
395
      block >>= 2;
      xoff = 8 * (block & twmask);
      yoff = 8 * (block >> twl);
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
396
      vp9_fdct8x8(src_diff, coeff, bw * 4);
397
398
399
      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);
400
401
      break;
    case TX_4X4:
402
403
      scan = vp9_default_scan_4x4;
      iscan = vp9_default_iscan_4x4;
404
405
406
      xoff = 4 * (block & twmask);
      yoff = 4 * (block >> twl);
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
407
      x->fwd_txm4x4(src_diff, coeff, bw * 4);
408
409
410
      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);
411
412
413
      break;
    default:
      assert(0);
John Koleszar's avatar
John Koleszar committed
414
  }
415
416
}

417
static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize,
418
                         TX_SIZE tx_size, void *arg) {
419
420
421
  struct encode_b_args *const args = arg;
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
422
  struct optimize_ctx *const ctx = args->ctx;
423
  struct macroblockd_plane *const pd = &xd->plane[plane];
424
  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
425
426
427
428
  int i, j;
  uint8_t *dst;
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
  dst = &pd->dst.buf[4 * j * pd->dst.stride + 4 * i];
429
430
431
432
433

  // TODO(jingning): per transformed block zero forcing only enabled for
  // luma component. will integrate chroma components as well.
  if (x->zcoeff_blk[tx_size][block] && plane == 0) {
    pd->eobs[block] = 0;
434
    ctx->ta[plane][i] = 0;
435
    ctx->tl[plane][j] = 0;
436
437
438
    return;
  }

439
  if (!x->skip_recode)
440
    vp9_xform_quant(plane, block, plane_bsize, tx_size, arg);
441

442
  if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
443
    vp9_optimize_b(plane, block, plane_bsize, tx_size, x, ctx);
444
445
  } else {
    ctx->ta[plane][i] = pd->eobs[block] > 0;
446
    ctx->tl[plane][j] = pd->eobs[block] > 0;
447
  }
John Koleszar's avatar
John Koleszar committed
448

449
  if (x->skip_encode || pd->eobs[block] == 0)
450
    return;
451

452
  switch (tx_size) {
453
    case TX_32X32:
Yaowu Xu's avatar
Yaowu Xu committed
454
      vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
455
456
      break;
    case TX_16X16:
457
      vp9_idct16x16_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
458
459
      break;
    case TX_8X8:
460
      vp9_idct8x8_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
461
462
      break;
    case TX_4X4:
463
464
465
      // 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.
466
      xd->itxm_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
467
      break;
468
469
    default:
      assert(!"Invalid transform size");
470
  }
John Koleszar's avatar
John Koleszar committed
471
472
}

473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
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]);
}

494
void vp9_encode_sby(MACROBLOCK *x, BLOCK_SIZE bsize) {
495
  MACROBLOCKD *const xd = &x->e_mbd;
496
  struct optimize_ctx ctx;
497
  struct encode_b_args arg = {x, &ctx};
498
499
500

  vp9_subtract_sby(x, bsize);
  if (x->optimize)
501
    optimize_init_b(0, bsize, &arg);
502

503
  foreach_transformed_block_in_plane(xd, bsize, 0, encode_block_pass1, &arg);
504
505
}

506
void vp9_encode_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
507
  MACROBLOCKD *const xd = &x->e_mbd;
508
  struct optimize_ctx ctx;
509
  struct encode_b_args arg = {x, &ctx};
John Koleszar's avatar
John Koleszar committed
510

511
  if (!x->skip_recode)
512
    vp9_subtract_sb(x, bsize);
513

514
  if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
515
516
517
518
    int i;
    for (i = 0; i < MAX_MB_PLANE; ++i)
      optimize_init_b(i, bsize, &arg);
  }
Scott LaVarnway's avatar
Scott LaVarnway committed
519

Scott LaVarnway's avatar
Scott LaVarnway committed
520
  foreach_transformed_block(xd, bsize, encode_block, &arg);
John Koleszar's avatar
John Koleszar committed
521
}
522

523
void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
524
                            TX_SIZE tx_size, void *arg) {
525
  struct encode_b_args* const args = arg;
526
527
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
528
  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
529
530
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
531
532
533
  int16_t *coeff = BLOCK_OFFSET(p->coeff, block);
  int16_t *qcoeff = BLOCK_OFFSET(pd->qcoeff, block);
  int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
534
  const int16_t *scan, *iscan;
535
  TX_TYPE tx_type;
536
  MB_PREDICTION_MODE mode;
537
  const int bwl = b_width_log2(plane_bsize), bw = 1 << bwl;
538
539
540
541
542
  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];
543

544
  if (xd->mb_to_right_edge < 0 || xd->mb_to_bottom_edge < 0)
545
    extend_for_intra(xd, plane_bsize, plane, block, tx_size);
546

547
  // if (x->optimize)
548
  // vp9_optimize_b(plane, block, plane_bsize, tx_size, x, args->ctx);
549

550
  switch (tx_size) {
551
    case TX_32X32:
552
553
554
555
556
557
558
559
560
      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);
561
562
563
564
565
566
567
568
569
570
571
572
573
574

      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);
      }
575
      if (!x->skip_encode && *eob)
Yaowu Xu's avatar
Yaowu Xu committed
576
        vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, *eob);
577
578
      break;
    case TX_16X16:
579
      tx_type = get_tx_type_16x16(pd->plane_type, xd);
580
581
582
583
584
585
586
587
588
      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);
589
590
591
592
593
594
595
596
597
598
      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);
      }
599
      if (!x->skip_encode && *eob)
600
        vp9_iht16x16_add(tx_type, dqcoeff, dst, pd->dst.stride, *eob);
601
602
      break;
    case TX_8X8:
603
      tx_type = get_tx_type_8x8(pd->plane_type, xd);
604
605
606
607
608
609
610
611
612
      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);
613
614
615
616
617
618
619
620
621
622
      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);
      }
623
      if (!x->skip_encode && *eob)
624
        vp9_iht8x8_add(tx_type, dqcoeff, dst, pd->dst.stride, *eob);
625
626
      break;
    case TX_4X4:
627
      tx_type = get_tx_type_4x4(pd->plane_type, xd, block);
628
629
      scan = get_scan_4x4(tx_type);
      iscan = get_iscan_4x4(tx_type);
630
      if (mbmi->sb_type < BLOCK_8X8 && plane == 0)
631
        mode = xd->mi_8x8[0]->bmi[block].as_mode;
632
      else
633
        mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
634

635
636
637
638
639
      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);
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654

      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);
      }

655
      if (!x->skip_encode && *eob) {
656
657
658
659
        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.
660
          xd->itxm_add(dqcoeff, dst, pd->dst.stride, *eob);
661
        else
662
          vp9_iht4x4_16_add(dqcoeff, dst, pd->dst.stride, tx_type);
663
      }
664
      break;
665
666
    default:
      assert(0);
667
668
669
  }
}

670
void vp9_encode_intra_block_y(MACROBLOCK *x, BLOCK_SIZE bsize) {
671
672
  MACROBLOCKD* const xd = &x->e_mbd;
  struct optimize_ctx ctx;
673
  struct encode_b_args arg = {x, &ctx};
674

675
676
  foreach_transformed_block_in_plane(xd, bsize, 0, vp9_encode_block_intra,
                                     &arg);
677
}
678
void vp9_encode_intra_block_uv(MACROBLOCK *x, BLOCK_SIZE bsize) {
679
680
  MACROBLOCKD* const xd = &x->e_mbd;
  struct optimize_ctx ctx;
681
  struct encode_b_args arg = {x, &ctx};
682
  foreach_transformed_block_uv(xd, bsize, vp9_encode_block_intra, &arg);
683
}
684