vp9_encodemb.c 24.5 KB
Newer Older
John Koleszar's avatar
John Koleszar committed
1
/*
2
 *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
John Koleszar's avatar
John Koleszar committed
3
 *
4
 *  Use of this source code is governed by a BSD-style license
5
6
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
7
 *  in the file PATENTS.  All contributing project authors may
8
 *  be found in the AUTHORS file in the root of the source tree.
John Koleszar's avatar
John Koleszar committed
9
10
 */

11
12

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

#include "vpx_mem/vpx_mem.h"

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

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

28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
void vp9_setup_interp_filters(MACROBLOCKD *xd,
                              INTERPOLATION_TYPE mcomp_filter_type,
                              VP9_COMMON *cm) {
  if (xd->mi_8x8 && xd->mi_8x8[0]) {
    MB_MODE_INFO *const mbmi = &xd->mi_8x8[0]->mbmi;

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

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

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

48
49
50
51
void vp9_subtract_block_c(int rows, int cols,
                          int16_t *diff_ptr, ptrdiff_t diff_stride,
                          const uint8_t *src_ptr, ptrdiff_t src_stride,
                          const uint8_t *pred_ptr, ptrdiff_t pred_stride) {
John Koleszar's avatar
John Koleszar committed
52
  int r, c;
John Koleszar's avatar
John Koleszar committed
53

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

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

64
static void subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
65
66
67
68
69
70
71
72
73
  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
74
75
}

76
void vp9_subtract_sby(MACROBLOCK *x, BLOCK_SIZE bsize) {
77
  subtract_plane(x, bsize, 0);
78
79
}

80
void vp9_subtract_sbuv(MACROBLOCK *x, BLOCK_SIZE bsize) {
81
  int i;
82

83
84
  for (i = 1; i < MAX_MB_PLANE; i++)
    subtract_plane(x, bsize, i);
85
86
}

87
void vp9_subtract_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
88
89
  vp9_subtract_sby(x, bsize);
  vp9_subtract_sbuv(x, bsize);
John Koleszar's avatar
John Koleszar committed
90
91
}

92
#define RDTRUNC(RM, DM, R, D) ((128 + (R) * (RM)) & 0xFF)
93
typedef struct vp9_token_state vp9_token_state;
94

95
struct vp9_token_state {
96
97
  int           rate;
  int           error;
Daniel Kang's avatar
Daniel Kang committed
98
  int           next;
99
100
101
102
  signed char   token;
  short         qc;
};

103
// TODO(jimbankoski): experiment to find optimal RD numbers.
104
105
#define Y1_RD_MULT 4
#define UV_RD_MULT 2
106

John Koleszar's avatar
John Koleszar committed
107
108
109
static const int plane_rd_mult[4] = {
  Y1_RD_MULT,
  UV_RD_MULT,
110
111
};

112
113
114
115
116
117
118
119
120
121
#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);\
  }\
}

122
123
// This function is a place holder for now but may ultimately need
// to scan previous tokens to work out the correct context.
124
125
static int trellis_get_coeff_context(const int16_t *scan,
                                     const int16_t *nb,
126
                                     int idx, int token,
127
                                     uint8_t *token_cache) {
128
  int bak = token_cache[scan[idx]], pt;
129
  token_cache[scan[idx]] = vp9_pt_energy_class[token];
130
  pt = get_coef_context(nb, token_cache, idx + 1);
131
  token_cache[scan[idx]] = bak;
132
  return pt;
133
134
}

135
static void optimize_b(MACROBLOCK *mb,
136
                       int plane, int block, BLOCK_SIZE plane_bsize,
137
                       ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
John Koleszar's avatar
John Koleszar committed
138
                       TX_SIZE tx_size) {
139
  MACROBLOCKD *const xd = &mb->e_mbd;
140
  struct macroblockd_plane *pd = &xd->plane[plane];
141
  const int ref = is_inter_block(&xd->mi_8x8[0]->mbmi);
142
143
  vp9_token_state tokens[1025][2];
  unsigned best_index[1025][2];
144
  const int16_t *coeff_ptr = BLOCK_OFFSET(mb->plane[plane].coeff, block);
145
146
  int16_t *qcoeff_ptr;
  int16_t *dqcoeff_ptr;
147
  int eob = pd->eobs[block], final_eob, sz = 0;
148
  const int i0 = 0;
149
  int rc, x, next, i;
150
151
152
  int64_t rdmult, rddiv, rd_cost0, rd_cost1;
  int rate0, rate1, error0, error1, t0, t1;
  int best, band, pt;
153
  PLANE_TYPE type = pd->plane_type;
John Koleszar's avatar
John Koleszar committed
154
  int err_mult = plane_rd_mult[type];
155
  const int default_eob = 16 << (tx_size << 1);
156
  const int16_t *scan, *nb;
157
  const int mul = 1 + (tx_size == TX_32X32);
158
  uint8_t token_cache[1024];
159
  const int16_t *dequant_ptr = pd->dequant;
160
  const uint8_t *const band_translate = get_band_translate(tx_size);
John Koleszar's avatar
John Koleszar committed
161

John Koleszar's avatar
John Koleszar committed
162
  assert((!type && !plane) || (type && plane));
163
164
  dqcoeff_ptr = BLOCK_OFFSET(pd->dqcoeff, block);
  qcoeff_ptr = BLOCK_OFFSET(pd->qcoeff, block);
165
  get_scan(xd, tx_size, type, block, &scan, &nb);
John Koleszar's avatar
John Koleszar committed
166
  assert(eob <= default_eob);
John Koleszar's avatar
John Koleszar committed
167
168
169

  /* Now set up a Viterbi trellis to evaluate alternative roundings. */
  rdmult = mb->rdmult * err_mult;
170
  if (!is_inter_block(&mb->e_mbd.mi_8x8[0]->mbmi))
John Koleszar's avatar
John Koleszar committed
171
172
173
174
175
    rdmult = (rdmult * 9) >> 4;
  rddiv = mb->rddiv;
  /* Initialize the sentinel node of the trellis. */
  tokens[eob][0].rate = 0;
  tokens[eob][0].error = 0;
176
  tokens[eob][0].next = default_eob;
John Koleszar's avatar
John Koleszar committed
177
178
179
180
  tokens[eob][0].token = DCT_EOB_TOKEN;
  tokens[eob][0].qc = 0;
  *(tokens[eob] + 1) = *(tokens[eob] + 0);
  next = eob;
181
  for (i = 0; i < eob; i++)
182
183
    token_cache[scan[i]] = vp9_pt_energy_class[vp9_dct_value_tokens_ptr[
        qcoeff_ptr[scan[i]]].token];
184

John Koleszar's avatar
John Koleszar committed
185
  for (i = eob; i-- > i0;) {
186
    int base_bits, d2, dx;
John Koleszar's avatar
John Koleszar committed
187

188
    rc = scan[i];
John Koleszar's avatar
John Koleszar committed
189
190
191
192
193
194
195
196
197
    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;
198
      t0 = (vp9_dct_value_tokens_ptr + x)->token;
John Koleszar's avatar
John Koleszar committed
199
      /* Consider both possible successor states. */
200
      if (next < default_eob) {
201
        band = band_translate[i + 1];
202
        pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
John Koleszar's avatar
John Koleszar committed
203
        rate0 +=
204
          mb->token_costs[tx_size][type][ref][band][0][pt]
205
                         [tokens[next][0].token];
John Koleszar's avatar
John Koleszar committed
206
        rate1 +=
207
          mb->token_costs[tx_size][type][ref][band][0][pt]
208
                         [tokens[next][1].token];
John Koleszar's avatar
John Koleszar committed
209
      }
210
      UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
211
212
      /* And pick the best. */
      best = rd_cost1 < rd_cost0;
213
      base_bits = *(vp9_dct_value_cost_ptr + x);
214
      dx = mul * (dqcoeff_ptr[rc] - coeff_ptr[rc]);
John Koleszar's avatar
John Koleszar committed
215
216
217
218
219
220
      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;
221
      best_index[i][0] = best;
222

John Koleszar's avatar
John Koleszar committed
223
224
225
226
      /* Evaluate the second possibility for this state. */
      rate0 = tokens[next][0].rate;
      rate1 = tokens[next][1].rate;

227
228
229
      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
230
231
232
233
234
235
236
237
238
239
240
241
242
        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.
243
         */
John Koleszar's avatar
John Koleszar committed
244
245
246
247
248
        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 {
249
        t0 = t1 = (vp9_dct_value_tokens_ptr + x)->token;
John Koleszar's avatar
John Koleszar committed
250
      }
251
      if (next < default_eob) {
252
        band = band_translate[i + 1];
John Koleszar's avatar
John Koleszar committed
253
        if (t0 != DCT_EOB_TOKEN) {
254
          pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
255
          rate0 += mb->token_costs[tx_size][type][ref][band][!x][pt]
256
                                  [tokens[next][0].token];
John Koleszar's avatar
John Koleszar committed
257
        }
John Koleszar's avatar
John Koleszar committed
258
        if (t1 != DCT_EOB_TOKEN) {
259
          pt = trellis_get_coeff_context(scan, nb, i, t1, token_cache);
260
          rate1 += mb->token_costs[tx_size][type][ref][band][!x][pt]
261
                                  [tokens[next][1].token];
John Koleszar's avatar
John Koleszar committed
262
263
        }
      }
John Koleszar's avatar
John Koleszar committed
264

265
      UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
266
267
      /* And pick the best. */
      best = rd_cost1 < rd_cost0;
268
      base_bits = *(vp9_dct_value_cost_ptr + x);
John Koleszar's avatar
John Koleszar committed
269
270

      if (shortcut) {
271
        dx -= (dequant_ptr[rc != 0] + sz) ^ sz;
John Koleszar's avatar
John Koleszar committed
272
273
274
275
276
277
278
        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;
279
      best_index[i][1] = best;
John Koleszar's avatar
John Koleszar committed
280
281
      /* Finally, make this the new head of the trellis. */
      next = i;
282
283
284
285
    } 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.
       */
286
      band = band_translate[i + 1];
John Koleszar's avatar
John Koleszar committed
287
288
289
290
      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) {
291
        tokens[next][0].rate +=
292
            mb->token_costs[tx_size][type][ref][band][1][0][t0];
John Koleszar's avatar
John Koleszar committed
293
294
295
        tokens[next][0].token = ZERO_TOKEN;
      }
      if (t1 != DCT_EOB_TOKEN) {
296
        tokens[next][1].rate +=
297
            mb->token_costs[tx_size][type][ref][band][1][0][t1];
John Koleszar's avatar
John Koleszar committed
298
299
        tokens[next][1].token = ZERO_TOKEN;
      }
300
      best_index[i][0] = best_index[i][1] = 0;
John Koleszar's avatar
John Koleszar committed
301
      /* Don't update next, because we didn't add a new node. */
302
    }
John Koleszar's avatar
John Koleszar committed
303
304
305
  }

  /* Now pick the best path through the whole trellis. */
306
  band = band_translate[i + 1];
307
  pt = combine_entropy_contexts(*a, *l);
John Koleszar's avatar
John Koleszar committed
308
309
310
311
312
313
  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;
314
315
  rate0 += mb->token_costs[tx_size][type][ref][band][0][pt][t0];
  rate1 += mb->token_costs[tx_size][type][ref][band][0][pt][t1];
316
  UPDATE_RD_COST();
John Koleszar's avatar
John Koleszar committed
317
318
  best = rd_cost1 < rd_cost0;
  final_eob = i0 - 1;
319
320
  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
321
322
  for (i = next; i < eob; i = next) {
    x = tokens[i][best].qc;
323
    if (x) {
John Koleszar's avatar
John Koleszar committed
324
      final_eob = i;
325
    }
326
    rc = scan[i];
John Koleszar's avatar
John Koleszar committed
327
    qcoeff_ptr[rc] = x;
328
    dqcoeff_ptr[rc] = (x * dequant_ptr[rc != 0]) / mul;
329

John Koleszar's avatar
John Koleszar committed
330
    next = tokens[i][best].next;
331
    best = best_index[i][best];
John Koleszar's avatar
John Koleszar committed
332
333
334
  }
  final_eob++;

John Koleszar's avatar
John Koleszar committed
335
  xd->plane[plane].eobs[block] = final_eob;
336
  *a = *l = (final_eob > 0);
337
338
}

339
void vp9_optimize_b(int plane, int block, BLOCK_SIZE plane_bsize,
340
                    TX_SIZE tx_size, MACROBLOCK *mb, struct optimize_ctx *ctx) {
John Koleszar's avatar
John Koleszar committed
341
  int x, y;
342
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x, &y);
343
344
  optimize_b(mb, plane, block, plane_bsize,
             &ctx->ta[plane][x], &ctx->tl[plane][y], tx_size);
345
346
}

347
static void optimize_init_b(int plane, BLOCK_SIZE bsize,
348
                            struct encode_b_args *args) {
349
350
  const MACROBLOCKD *xd = &args->x->e_mbd;
  const struct macroblockd_plane* const pd = &xd->plane[plane];
351
  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
352
353
  const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
  const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
354
  const MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
355
  const TX_SIZE tx_size = plane ? get_uv_tx_size(mbmi) : mbmi->tx_size;
356

357
358
359
  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);
360
361
}

362
void vp9_xform_quant(int plane, int block, BLOCK_SIZE plane_bsize,
363
                     TX_SIZE tx_size, void *arg) {
364
365
366
  struct encode_b_args* const args = arg;
  MACROBLOCK* const x = args->x;
  MACROBLOCKD* const xd = &x->e_mbd;
367
368
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
369
370
371
  int16_t *coeff = BLOCK_OFFSET(p->coeff, block);
  int16_t *qcoeff = BLOCK_OFFSET(pd->qcoeff, block);
  int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
372
373
  const int16_t *scan, *iscan;
  uint16_t *eob = &pd->eobs[block];
374
  const int bwl = b_width_log2(plane_bsize), bw = 1 << bwl;
375
376
377
378
379
  const int twl = bwl - tx_size, twmask = (1 << twl) - 1;
  int xoff, yoff;
  int16_t *src_diff;

  switch (tx_size) {
380
    case TX_32X32:
381
382
383
384
385
386
      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;
387
      if (x->use_lp32x32fdct)
388
        vp9_fdct32x32_rd(src_diff, coeff, bw * 4);
389
      else
390
        vp9_fdct32x32(src_diff, coeff, bw * 4);
391
392
393
      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);
394
395
      break;
    case TX_16X16:
396
397
      scan = vp9_default_scan_16x16;
      iscan = vp9_default_iscan_16x16;
398
399
400
401
      block >>= 4;
      xoff = 16 * (block & twmask);
      yoff = 16 * (block >> twl);
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
402
      vp9_fdct16x16(src_diff, coeff, bw * 4);
403
404
405
      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);
406
407
      break;
    case TX_8X8:
408
409
      scan = vp9_default_scan_8x8;
      iscan = vp9_default_iscan_8x8;
410
411
412
413
      block >>= 2;
      xoff = 8 * (block & twmask);
      yoff = 8 * (block >> twl);
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
414
      vp9_fdct8x8(src_diff, coeff, bw * 4);
415
416
417
      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);
418
419
      break;
    case TX_4X4:
420
421
      scan = vp9_default_scan_4x4;
      iscan = vp9_default_iscan_4x4;
422
423
424
      xoff = 4 * (block & twmask);
      yoff = 4 * (block >> twl);
      src_diff = p->src_diff + 4 * bw * yoff + xoff;
425
      x->fwd_txm4x4(src_diff, coeff, bw * 4);
426
427
428
      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);
429
430
431
      break;
    default:
      assert(0);
John Koleszar's avatar
John Koleszar committed
432
  }
433
434
}

435
static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize,
436
                         TX_SIZE tx_size, void *arg) {
437
438
439
  struct encode_b_args *const args = arg;
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
440
  struct optimize_ctx *const ctx = args->ctx;
441
  struct macroblockd_plane *const pd = &xd->plane[plane];
442
  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
443
444
445
446
  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];
447
448
449
450
451

  // 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;
452
    ctx->ta[plane][i] = 0;
453
    ctx->tl[plane][j] = 0;
454
455
456
    return;
  }

457
  if (!x->skip_recode)
458
    vp9_xform_quant(plane, block, plane_bsize, tx_size, arg);
459

460
  if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
461
    vp9_optimize_b(plane, block, plane_bsize, tx_size, x, ctx);
462
463
  } else {
    ctx->ta[plane][i] = pd->eobs[block] > 0;
464
    ctx->tl[plane][j] = pd->eobs[block] > 0;
465
  }
John Koleszar's avatar
John Koleszar committed
466

467
  if (x->skip_encode || pd->eobs[block] == 0)
468
    return;
469

470
  switch (tx_size) {
471
    case TX_32X32:
Yaowu Xu's avatar
Yaowu Xu committed
472
      vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
473
474
      break;
    case TX_16X16:
475
      vp9_idct16x16_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
476
477
      break;
    case TX_8X8:
478
      vp9_idct8x8_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
479
480
      break;
    case TX_4X4:
481
482
483
      // 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.
484
      xd->itxm_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
485
      break;
486
487
    default:
      assert(!"Invalid transform size");
488
  }
John Koleszar's avatar
John Koleszar committed
489
490
}

491
492
493
494
495
496
497
static void encode_block_pass1(int plane, int block, BLOCK_SIZE plane_bsize,
                               TX_SIZE tx_size, void *arg) {
  struct encode_b_args *const args = arg;
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
  struct macroblockd_plane *const pd = &xd->plane[plane];
  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
498
499
500
501
  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];
502
503
504
505
506
507
508
509
510

  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]);
}

511
void vp9_encode_sby(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};
515
516
517

  vp9_subtract_sby(x, bsize);
  if (x->optimize)
518
    optimize_init_b(0, bsize, &arg);
519

520
  foreach_transformed_block_in_plane(xd, bsize, 0, encode_block_pass1, &arg);
521
522
}

523
void vp9_encode_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
524
  MACROBLOCKD *const xd = &x->e_mbd;
525
  struct optimize_ctx ctx;
526
  struct encode_b_args arg = {x, &ctx};
John Koleszar's avatar
John Koleszar committed
527

528
  if (!x->skip_recode)
529
    vp9_subtract_sb(x, bsize);
530

531
  if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
532
533
534
535
    int i;
    for (i = 0; i < MAX_MB_PLANE; ++i)
      optimize_init_b(i, bsize, &arg);
  }
Scott LaVarnway's avatar
Scott LaVarnway committed
536

Scott LaVarnway's avatar
Scott LaVarnway committed
537
  foreach_transformed_block(xd, bsize, encode_block, &arg);
John Koleszar's avatar
John Koleszar committed
538
}
539

540
void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
541
                            TX_SIZE tx_size, void *arg) {
542
  struct encode_b_args* const args = arg;
543
544
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
545
  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
546
547
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
548
549
550
  int16_t *coeff = BLOCK_OFFSET(p->coeff, block);
  int16_t *qcoeff = BLOCK_OFFSET(pd->qcoeff, block);
  int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
551
  const int16_t *scan, *iscan;
552
  TX_TYPE tx_type;
553
  MB_PREDICTION_MODE mode;
554
555
  const int bwl = b_width_log2(plane_bsize);
  const int diff_stride = 4 * (1 << bwl);
556
557
558
  uint8_t *src, *dst;
  int16_t *src_diff;
  uint16_t *eob = &pd->eobs[block];
559
560
561
562
563
  int i, j;
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
  dst = &pd->dst.buf[4 * (j * pd->dst.stride + i)];
  src = &p->src.buf[4 * (j * p->src.stride + i)];
  src_diff = &p->src_diff[4 * (j * diff_stride + i)];
564

565
  if (xd->mb_to_right_edge < 0 || xd->mb_to_bottom_edge < 0)
566
    extend_for_intra(xd, plane_bsize, plane, i, j);
567

568
  // if (x->optimize)
569
  // vp9_optimize_b(plane, block, plane_bsize, tx_size, x, args->ctx);
570

571
  switch (tx_size) {
572
    case TX_32X32:
573
574
575
576
577
578
      scan = vp9_default_scan_32x32;
      iscan = vp9_default_iscan_32x32;
      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
      block >>= 6;
      vp9_predict_intra_block(xd, block, bwl, TX_32X32, mode,
                              dst, pd->dst.stride, dst, pd->dst.stride);
579
580

      if (!x->skip_recode) {
581
        vp9_subtract_block(32, 32, src_diff, diff_stride,
582
583
                           src, p->src.stride, dst, pd->dst.stride);
        if (x->use_lp32x32fdct)
584
          vp9_fdct32x32_rd(src_diff, coeff, diff_stride);
585
        else
586
          vp9_fdct32x32(src_diff, coeff, diff_stride);
587
588
589
590
        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);
      }
591
      if (!x->skip_encode && *eob)
Yaowu Xu's avatar
Yaowu Xu committed
592
        vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, *eob);
593
594
      break;
    case TX_16X16:
595
      tx_type = get_tx_type_16x16(pd->plane_type, xd);
596
597
598
599
600
601
      scan = get_scan_16x16(tx_type);
      iscan = get_iscan_16x16(tx_type);
      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
      block >>= 4;
      vp9_predict_intra_block(xd, block, bwl, TX_16X16, mode,
                              dst, pd->dst.stride, dst, pd->dst.stride);
602
      if (!x->skip_recode) {
603
        vp9_subtract_block(16, 16, src_diff, diff_stride,
604
                           src, p->src.stride, dst, pd->dst.stride);
605
        vp9_fht16x16(tx_type, src_diff, coeff, diff_stride);
606
607
608
609
        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);
      }
610
      if (!x->skip_encode && *eob)
611
        vp9_iht16x16_add(tx_type, dqcoeff, dst, pd->dst.stride, *eob);
612
613
      break;
    case TX_8X8:
614
      tx_type = get_tx_type_8x8(pd->plane_type, xd);
615
616
617
618
619
620
      scan = get_scan_8x8(tx_type);
      iscan = get_iscan_8x8(tx_type);
      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
      block >>= 2;
      vp9_predict_intra_block(xd, block, bwl, TX_8X8, mode,
                              dst, pd->dst.stride, dst, pd->dst.stride);
621
      if (!x->skip_recode) {
622
        vp9_subtract_block(8, 8, src_diff, diff_stride,
623
                           src, p->src.stride, dst, pd->dst.stride);
624
        vp9_fht8x8(tx_type, src_diff, coeff, diff_stride);
625
626
627
628
        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
      vp9_predict_intra_block(xd, block, bwl, TX_4X4, mode,
                              dst, pd->dst.stride, dst, pd->dst.stride);
643
644

      if (!x->skip_recode) {
645
        vp9_subtract_block(4, 4, src_diff, diff_stride,
646
647
                           src, p->src.stride, dst, pd->dst.stride);
        if (tx_type != DCT_DCT)
648
          vp9_short_fht4x4(src_diff, coeff, diff_stride, tx_type);
649
        else
650
          x->fwd_txm4x4(src_diff, coeff, diff_stride);
651
652
653
654
655
        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);
      }

656
      if (!x->skip_encode && *eob) {
657
658
659
660
        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.
661
          xd->itxm_add(dqcoeff, dst, pd->dst.stride, *eob);
662
        else
663
          vp9_iht4x4_16_add(dqcoeff, dst, pd->dst.stride, tx_type);
664
      }
665
      break;
666
667
    default:
      assert(0);
668
669
670
  }
}

671
void vp9_encode_intra_block_y(MACROBLOCK *x, BLOCK_SIZE bsize) {
672
673
  MACROBLOCKD* const xd = &x->e_mbd;
  struct optimize_ctx ctx;
674
  struct encode_b_args arg = {x, &ctx};
675

676
677
  foreach_transformed_block_in_plane(xd, bsize, 0, vp9_encode_block_intra,
                                     &arg);
678
}
679
void vp9_encode_intra_block_uv(MACROBLOCK *x, BLOCK_SIZE bsize) {
680
681
  MACROBLOCKD* const xd = &x->e_mbd;
  struct optimize_ctx ctx;
682
  struct encode_b_args arg = {x, &ctx};
683
  foreach_transformed_block_uv(xd, bsize, vp9_encode_block_intra, &arg);
684
}
685

686
687
688
689
690
691
692
693
694
695
696
int vp9_encode_intra(MACROBLOCK *x, int use_16x16_pred) {
  MB_MODE_INFO * mbmi = &x->e_mbd.mi_8x8[0]->mbmi;
  x->skip_encode = 0;
  mbmi->mode = DC_PRED;
  mbmi->ref_frame[0] = INTRA_FRAME;
  mbmi->tx_size = use_16x16_pred ? (mbmi->sb_type >= BLOCK_16X16 ? TX_16X16
                                                                 : TX_8X8)
                                   : TX_4X4;
  vp9_encode_intra_block_y(x, mbmi->sb_type);
  return vp9_get_mb_ss(x->plane[0].src_diff);
}