1/*
2 *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
3 *
4 *  Use of this source code is governed by a BSD-style license
5 *  that can be found in the LICENSE file in the root of the source
6 *  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.
9 */
10
11
12#include "./vp9_rtcd.h"
13#include "./vpx_config.h"
14
15#include "vpx_mem/vpx_mem.h"
16
17#include "vp9/common/vp9_idct.h"
18#include "vp9/common/vp9_reconinter.h"
19#include "vp9/common/vp9_reconintra.h"
20#include "vp9/common/vp9_systemdependent.h"
21
22#include "vp9/encoder/vp9_dct.h"
23#include "vp9/encoder/vp9_encodemb.h"
24#include "vp9/encoder/vp9_quantize.h"
25#include "vp9/encoder/vp9_rdopt.h"
26#include "vp9/encoder/vp9_tokenize.h"
27
28void vp9_subtract_block_c(int rows, int cols,
29                          int16_t *diff_ptr, ptrdiff_t diff_stride,
30                          const uint8_t *src_ptr, ptrdiff_t src_stride,
31                          const uint8_t *pred_ptr, ptrdiff_t pred_stride) {
32  int r, c;
33
34  for (r = 0; r < rows; r++) {
35    for (c = 0; c < cols; c++)
36      diff_ptr[c] = src_ptr[c] - pred_ptr[c];
37
38    diff_ptr += diff_stride;
39    pred_ptr += pred_stride;
40    src_ptr  += src_stride;
41  }
42}
43
44static void subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
45  struct macroblock_plane *const p = &x->plane[plane];
46  const MACROBLOCKD *const xd = &x->e_mbd;
47  const struct macroblockd_plane *const pd = &xd->plane[plane];
48  const int bw = plane_block_width(bsize, pd);
49  const int bh = plane_block_height(bsize, pd);
50
51  vp9_subtract_block(bh, bw, p->src_diff, bw,
52                     p->src.buf, p->src.stride,
53                     pd->dst.buf, pd->dst.stride);
54}
55
56void vp9_subtract_sby(MACROBLOCK *x, BLOCK_SIZE bsize) {
57  subtract_plane(x, bsize, 0);
58}
59
60void vp9_subtract_sbuv(MACROBLOCK *x, BLOCK_SIZE bsize) {
61  int i;
62
63  for (i = 1; i < MAX_MB_PLANE; i++)
64    subtract_plane(x, bsize, i);
65}
66
67void vp9_subtract_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
68  vp9_subtract_sby(x, bsize);
69  vp9_subtract_sbuv(x, bsize);
70}
71
72#define RDTRUNC(RM, DM, R, D) ((128 + (R) * (RM)) & 0xFF)
73typedef struct vp9_token_state vp9_token_state;
74
75struct vp9_token_state {
76  int           rate;
77  int           error;
78  int           next;
79  signed char   token;
80  short         qc;
81};
82
83// TODO(jimbankoski): experiment to find optimal RD numbers.
84#define Y1_RD_MULT 4
85#define UV_RD_MULT 2
86
87static const int plane_rd_mult[4] = {
88  Y1_RD_MULT,
89  UV_RD_MULT,
90};
91
92#define UPDATE_RD_COST()\
93{\
94  rd_cost0 = RDCOST(rdmult, rddiv, rate0, error0);\
95  rd_cost1 = RDCOST(rdmult, rddiv, rate1, error1);\
96  if (rd_cost0 == rd_cost1) {\
97    rd_cost0 = RDTRUNC(rdmult, rddiv, rate0, error0);\
98    rd_cost1 = RDTRUNC(rdmult, rddiv, rate1, error1);\
99  }\
100}
101
102// This function is a place holder for now but may ultimately need
103// to scan previous tokens to work out the correct context.
104static int trellis_get_coeff_context(const int16_t *scan,
105                                     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
115static void optimize_b(MACROBLOCK *mb,
116                       int plane, int block, BLOCK_SIZE plane_bsize,
117                       ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
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  vp9_token_state tokens[1025][2];
123  unsigned best_index[1025][2];
124  const int16_t *coeff_ptr = BLOCK_OFFSET(mb->plane[plane].coeff, block);
125  int16_t *qcoeff_ptr;
126  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  int64_t rdmult, rddiv, rd_cost0, rd_cost1;
131  int rate0, rate1, error0, error1, t0, t1;
132  int best, band, pt;
133  PLANE_TYPE type = pd->plane_type;
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  const uint8_t *const band_translate = get_band_translate(tx_size);
141
142  assert((!type && !plane) || (type && plane));
143  dqcoeff_ptr = BLOCK_OFFSET(pd->dqcoeff, block);
144  qcoeff_ptr = BLOCK_OFFSET(pd->qcoeff, block);
145  get_scan(xd, tx_size, type, block, &scan, &nb);
146  assert(eob <= default_eob);
147
148  /* Now set up a Viterbi trellis to evaluate alternative roundings. */
149  rdmult = mb->rdmult * err_mult;
150  if (mb->e_mbd.mi_8x8[0]->mbmi.ref_frame[0] == INTRA_FRAME)
151    rdmult = (rdmult * 9) >> 4;
152  rddiv = mb->rddiv;
153  /* Initialize the sentinel node of the trellis. */
154  tokens[eob][0].rate = 0;
155  tokens[eob][0].error = 0;
156  tokens[eob][0].next = default_eob;
157  tokens[eob][0].token = DCT_EOB_TOKEN;
158  tokens[eob][0].qc = 0;
159  *(tokens[eob] + 1) = *(tokens[eob] + 0);
160  next = eob;
161  for (i = 0; i < eob; i++)
162    token_cache[scan[i]] = vp9_pt_energy_class[vp9_dct_value_tokens_ptr[
163        qcoeff_ptr[scan[i]]].token];
164
165  for (i = eob; i-- > i0;) {
166    int base_bits, d2, dx;
167
168    rc = scan[i];
169    x = qcoeff_ptr[rc];
170    /* Only add a trellis state for non-zero coefficients. */
171    if (x) {
172      int shortcut = 0;
173      error0 = tokens[next][0].error;
174      error1 = tokens[next][1].error;
175      /* Evaluate the first possibility for this state. */
176      rate0 = tokens[next][0].rate;
177      rate1 = tokens[next][1].rate;
178      t0 = (vp9_dct_value_tokens_ptr + x)->token;
179      /* Consider both possible successor states. */
180      if (next < default_eob) {
181        band = band_translate[i + 1];
182        pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
183        rate0 +=
184          mb->token_costs[tx_size][type][ref][band][0][pt]
185                         [tokens[next][0].token];
186        rate1 +=
187          mb->token_costs[tx_size][type][ref][band][0][pt]
188                         [tokens[next][1].token];
189      }
190      UPDATE_RD_COST();
191      /* And pick the best. */
192      best = rd_cost1 < rd_cost0;
193      base_bits = *(vp9_dct_value_cost_ptr + x);
194      dx = mul * (dqcoeff_ptr[rc] - coeff_ptr[rc]);
195      d2 = dx * dx;
196      tokens[i][0].rate = base_bits + (best ? rate1 : rate0);
197      tokens[i][0].error = d2 + (best ? error1 : error0);
198      tokens[i][0].next = next;
199      tokens[i][0].token = t0;
200      tokens[i][0].qc = x;
201      best_index[i][0] = best;
202
203      /* Evaluate the second possibility for this state. */
204      rate0 = tokens[next][0].rate;
205      rate1 = tokens[next][1].rate;
206
207      if ((abs(x)*dequant_ptr[rc != 0] > abs(coeff_ptr[rc]) * mul) &&
208          (abs(x)*dequant_ptr[rc != 0] < abs(coeff_ptr[rc]) * mul +
209                                         dequant_ptr[rc != 0]))
210        shortcut = 1;
211      else
212        shortcut = 0;
213
214      if (shortcut) {
215        sz = -(x < 0);
216        x -= 2 * sz + 1;
217      }
218
219      /* Consider both possible successor states. */
220      if (!x) {
221        /* If we reduced this coefficient to zero, check to see if
222         *  we need to move the EOB back here.
223         */
224        t0 = tokens[next][0].token == DCT_EOB_TOKEN ?
225             DCT_EOB_TOKEN : ZERO_TOKEN;
226        t1 = tokens[next][1].token == DCT_EOB_TOKEN ?
227             DCT_EOB_TOKEN : ZERO_TOKEN;
228      } else {
229        t0 = t1 = (vp9_dct_value_tokens_ptr + x)->token;
230      }
231      if (next < default_eob) {
232        band = band_translate[i + 1];
233        if (t0 != DCT_EOB_TOKEN) {
234          pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache);
235          rate0 += mb->token_costs[tx_size][type][ref][band][!x][pt]
236                                  [tokens[next][0].token];
237        }
238        if (t1 != DCT_EOB_TOKEN) {
239          pt = trellis_get_coeff_context(scan, nb, i, t1, token_cache);
240          rate1 += mb->token_costs[tx_size][type][ref][band][!x][pt]
241                                  [tokens[next][1].token];
242        }
243      }
244
245      UPDATE_RD_COST();
246      /* And pick the best. */
247      best = rd_cost1 < rd_cost0;
248      base_bits = *(vp9_dct_value_cost_ptr + x);
249
250      if (shortcut) {
251        dx -= (dequant_ptr[rc != 0] + sz) ^ sz;
252        d2 = dx * dx;
253      }
254      tokens[i][1].rate = base_bits + (best ? rate1 : rate0);
255      tokens[i][1].error = d2 + (best ? error1 : error0);
256      tokens[i][1].next = next;
257      tokens[i][1].token = best ? t1 : t0;
258      tokens[i][1].qc = x;
259      best_index[i][1] = best;
260      /* Finally, make this the new head of the trellis. */
261      next = i;
262    } else {
263      /* There's no choice to make for a zero coefficient, so we don't
264       *  add a new trellis node, but we do need to update the costs.
265       */
266      band = band_translate[i + 1];
267      t0 = tokens[next][0].token;
268      t1 = tokens[next][1].token;
269      /* Update the cost of each path if we're past the EOB token. */
270      if (t0 != DCT_EOB_TOKEN) {
271        tokens[next][0].rate +=
272            mb->token_costs[tx_size][type][ref][band][1][0][t0];
273        tokens[next][0].token = ZERO_TOKEN;
274      }
275      if (t1 != DCT_EOB_TOKEN) {
276        tokens[next][1].rate +=
277            mb->token_costs[tx_size][type][ref][band][1][0][t1];
278        tokens[next][1].token = ZERO_TOKEN;
279      }
280      best_index[i][0] = best_index[i][1] = 0;
281      /* Don't update next, because we didn't add a new node. */
282    }
283  }
284
285  /* Now pick the best path through the whole trellis. */
286  band = band_translate[i + 1];
287  pt = combine_entropy_contexts(*a, *l);
288  rate0 = tokens[next][0].rate;
289  rate1 = tokens[next][1].rate;
290  error0 = tokens[next][0].error;
291  error1 = tokens[next][1].error;
292  t0 = tokens[next][0].token;
293  t1 = tokens[next][1].token;
294  rate0 += mb->token_costs[tx_size][type][ref][band][0][pt][t0];
295  rate1 += mb->token_costs[tx_size][type][ref][band][0][pt][t1];
296  UPDATE_RD_COST();
297  best = rd_cost1 < rd_cost0;
298  final_eob = i0 - 1;
299  vpx_memset(qcoeff_ptr, 0, sizeof(*qcoeff_ptr) * (16 << (tx_size * 2)));
300  vpx_memset(dqcoeff_ptr, 0, sizeof(*dqcoeff_ptr) * (16 << (tx_size * 2)));
301  for (i = next; i < eob; i = next) {
302    x = tokens[i][best].qc;
303    if (x) {
304      final_eob = i;
305    }
306    rc = scan[i];
307    qcoeff_ptr[rc] = x;
308    dqcoeff_ptr[rc] = (x * dequant_ptr[rc != 0]) / mul;
309
310    next = tokens[i][best].next;
311    best = best_index[i][best];
312  }
313  final_eob++;
314
315  xd->plane[plane].eobs[block] = final_eob;
316  *a = *l = (final_eob > 0);
317}
318
319void vp9_optimize_b(int plane, int block, BLOCK_SIZE plane_bsize,
320                    TX_SIZE tx_size, MACROBLOCK *mb, struct optimize_ctx *ctx) {
321  int x, y;
322  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x, &y);
323  optimize_b(mb, plane, block, plane_bsize,
324             &ctx->ta[plane][x], &ctx->tl[plane][y], tx_size);
325}
326
327static void optimize_init_b(int plane, BLOCK_SIZE bsize,
328                            struct encode_b_args *args) {
329  const MACROBLOCKD *xd = &args->x->e_mbd;
330  const struct macroblockd_plane* const pd = &xd->plane[plane];
331  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
332  const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
333  const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
334  const MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
335  const TX_SIZE tx_size = plane ? get_uv_tx_size(mbmi) : mbmi->tx_size;
336
337  vp9_get_entropy_contexts(tx_size, args->ctx->ta[plane], args->ctx->tl[plane],
338                           pd->above_context, pd->left_context,
339                           num_4x4_w, num_4x4_h);
340}
341
342void vp9_xform_quant(int plane, int block, BLOCK_SIZE plane_bsize,
343                     TX_SIZE tx_size, void *arg) {
344  struct encode_b_args* const args = arg;
345  MACROBLOCK* const x = args->x;
346  MACROBLOCKD* const xd = &x->e_mbd;
347  struct macroblock_plane *const p = &x->plane[plane];
348  struct macroblockd_plane *const pd = &xd->plane[plane];
349  int16_t *coeff = BLOCK_OFFSET(p->coeff, block);
350  int16_t *qcoeff = BLOCK_OFFSET(pd->qcoeff, block);
351  int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
352  const int16_t *scan, *iscan;
353  uint16_t *eob = &pd->eobs[block];
354  const int bwl = b_width_log2(plane_bsize), bw = 1 << bwl;
355  const int twl = bwl - tx_size, twmask = (1 << twl) - 1;
356  int xoff, yoff;
357  int16_t *src_diff;
358
359  switch (tx_size) {
360    case TX_32X32:
361      scan = vp9_default_scan_32x32;
362      iscan = vp9_default_iscan_32x32;
363      block >>= 6;
364      xoff = 32 * (block & twmask);
365      yoff = 32 * (block >> twl);
366      src_diff = p->src_diff + 4 * bw * yoff + xoff;
367      if (x->use_lp32x32fdct)
368        vp9_fdct32x32_rd(src_diff, coeff, bw * 4);
369      else
370        vp9_fdct32x32(src_diff, coeff, bw * 4);
371      vp9_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round,
372                           p->quant, p->quant_shift, qcoeff, dqcoeff,
373                           pd->dequant, p->zbin_extra, eob, scan, iscan);
374      break;
375    case TX_16X16:
376      scan = vp9_default_scan_16x16;
377      iscan = vp9_default_iscan_16x16;
378      block >>= 4;
379      xoff = 16 * (block & twmask);
380      yoff = 16 * (block >> twl);
381      src_diff = p->src_diff + 4 * bw * yoff + xoff;
382      vp9_fdct16x16(src_diff, coeff, bw * 4);
383      vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
384                     p->quant, p->quant_shift, qcoeff, dqcoeff,
385                     pd->dequant, p->zbin_extra, eob, scan, iscan);
386      break;
387    case TX_8X8:
388      scan = vp9_default_scan_8x8;
389      iscan = vp9_default_iscan_8x8;
390      block >>= 2;
391      xoff = 8 * (block & twmask);
392      yoff = 8 * (block >> twl);
393      src_diff = p->src_diff + 4 * bw * yoff + xoff;
394      vp9_fdct8x8(src_diff, coeff, bw * 4);
395      vp9_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round,
396                     p->quant, p->quant_shift, qcoeff, dqcoeff,
397                     pd->dequant, p->zbin_extra, eob, scan, iscan);
398      break;
399    case TX_4X4:
400      scan = vp9_default_scan_4x4;
401      iscan = vp9_default_iscan_4x4;
402      xoff = 4 * (block & twmask);
403      yoff = 4 * (block >> twl);
404      src_diff = p->src_diff + 4 * bw * yoff + xoff;
405      x->fwd_txm4x4(src_diff, coeff, bw * 4);
406      vp9_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round,
407                     p->quant, p->quant_shift, qcoeff, dqcoeff,
408                     pd->dequant, p->zbin_extra, eob, scan, iscan);
409      break;
410    default:
411      assert(0);
412  }
413}
414
415static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize,
416                         TX_SIZE tx_size, void *arg) {
417  struct encode_b_args *const args = arg;
418  MACROBLOCK *const x = args->x;
419  MACROBLOCKD *const xd = &x->e_mbd;
420  struct optimize_ctx *const ctx = args->ctx;
421  struct macroblockd_plane *const pd = &xd->plane[plane];
422  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
423  int i, j;
424  uint8_t *dst;
425  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
426  dst = &pd->dst.buf[4 * j * pd->dst.stride + 4 * i];
427
428  // TODO(jingning): per transformed block zero forcing only enabled for
429  // luma component. will integrate chroma components as well.
430  if (x->zcoeff_blk[tx_size][block] && plane == 0) {
431    pd->eobs[block] = 0;
432    ctx->ta[plane][i] = 0;
433    ctx->tl[plane][j] = 0;
434    return;
435  }
436
437  if (!x->skip_recode)
438    vp9_xform_quant(plane, block, plane_bsize, tx_size, arg);
439
440  if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
441    vp9_optimize_b(plane, block, plane_bsize, tx_size, x, ctx);
442  } else {
443    ctx->ta[plane][i] = pd->eobs[block] > 0;
444    ctx->tl[plane][j] = pd->eobs[block] > 0;
445  }
446
447  if (x->skip_encode || pd->eobs[block] == 0)
448    return;
449
450  switch (tx_size) {
451    case TX_32X32:
452      vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
453      break;
454    case TX_16X16:
455      vp9_idct16x16_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
456      break;
457    case TX_8X8:
458      vp9_idct8x8_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
459      break;
460    case TX_4X4:
461      // this is like vp9_short_idct4x4 but has a special case around eob<=1
462      // which is significant (not just an optimization) for the lossless
463      // case.
464      xd->itxm_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
465      break;
466    default:
467      assert(!"Invalid transform size");
468  }
469}
470
471static void encode_block_pass1(int plane, int block, BLOCK_SIZE plane_bsize,
472                               TX_SIZE tx_size, void *arg) {
473  struct encode_b_args *const args = arg;
474  MACROBLOCK *const x = args->x;
475  MACROBLOCKD *const xd = &x->e_mbd;
476  struct macroblockd_plane *const pd = &xd->plane[plane];
477  const int raster_block = txfrm_block_to_raster_block(plane_bsize, tx_size,
478                                                       block);
479
480  int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
481  uint8_t *const dst = raster_block_offset_uint8(plane_bsize, raster_block,
482                                                 pd->dst.buf, pd->dst.stride);
483
484  vp9_xform_quant(plane, block, plane_bsize, tx_size, arg);
485
486  if (pd->eobs[block] == 0)
487    return;
488
489  xd->itxm_add(dqcoeff, dst, pd->dst.stride, pd->eobs[block]);
490}
491
492void vp9_encode_sby(MACROBLOCK *x, BLOCK_SIZE bsize) {
493  MACROBLOCKD *const xd = &x->e_mbd;
494  struct optimize_ctx ctx;
495  struct encode_b_args arg = {x, &ctx};
496
497  vp9_subtract_sby(x, bsize);
498  if (x->optimize)
499    optimize_init_b(0, bsize, &arg);
500
501  foreach_transformed_block_in_plane(xd, bsize, 0, encode_block_pass1, &arg);
502}
503
504void vp9_encode_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
505  MACROBLOCKD *const xd = &x->e_mbd;
506  struct optimize_ctx ctx;
507  struct encode_b_args arg = {x, &ctx};
508
509  if (!x->skip_recode)
510    vp9_subtract_sb(x, bsize);
511
512  if (x->optimize && (!x->skip_recode || !x->skip_optimize)) {
513    int i;
514    for (i = 0; i < MAX_MB_PLANE; ++i)
515      optimize_init_b(i, bsize, &arg);
516  }
517
518  foreach_transformed_block(xd, bsize, encode_block, &arg);
519}
520
521void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
522                            TX_SIZE tx_size, void *arg) {
523  struct encode_b_args* const args = arg;
524  MACROBLOCK *const x = args->x;
525  MACROBLOCKD *const xd = &x->e_mbd;
526  MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi;
527  struct macroblock_plane *const p = &x->plane[plane];
528  struct macroblockd_plane *const pd = &xd->plane[plane];
529  int16_t *coeff = BLOCK_OFFSET(p->coeff, block);
530  int16_t *qcoeff = BLOCK_OFFSET(pd->qcoeff, block);
531  int16_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
532  const int16_t *scan, *iscan;
533  TX_TYPE tx_type;
534  MB_PREDICTION_MODE mode;
535  const int bwl = b_width_log2(plane_bsize), bw = 1 << bwl;
536  const int twl = bwl - tx_size, twmask = (1 << twl) - 1;
537  int xoff, yoff;
538  uint8_t *src, *dst;
539  int16_t *src_diff;
540  uint16_t *eob = &pd->eobs[block];
541
542  if (xd->mb_to_right_edge < 0 || xd->mb_to_bottom_edge < 0)
543    extend_for_intra(xd, plane_bsize, plane, block, tx_size);
544
545  // if (x->optimize)
546  // vp9_optimize_b(plane, block, plane_bsize, tx_size, x, args->ctx);
547
548  switch (tx_size) {
549    case TX_32X32:
550      scan = vp9_default_scan_32x32;
551      iscan = vp9_default_iscan_32x32;
552      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
553      block >>= 6;
554      xoff = 32 * (block & twmask);
555      yoff = 32 * (block >> twl);
556      dst = pd->dst.buf + yoff * pd->dst.stride + xoff;
557      vp9_predict_intra_block(xd, block, bwl, TX_32X32, mode,
558                              dst, pd->dst.stride, dst, pd->dst.stride);
559
560      if (!x->skip_recode) {
561        src = p->src.buf + yoff * p->src.stride + xoff;
562        src_diff = p->src_diff + 4 * bw * yoff + xoff;
563        vp9_subtract_block(32, 32, src_diff, bw * 4,
564                           src, p->src.stride, dst, pd->dst.stride);
565        if (x->use_lp32x32fdct)
566          vp9_fdct32x32_rd(src_diff, coeff, bw * 4);
567        else
568          vp9_fdct32x32(src_diff, coeff, bw * 4);
569        vp9_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round,
570                             p->quant, p->quant_shift, qcoeff, dqcoeff,
571                             pd->dequant, p->zbin_extra, eob, scan, iscan);
572      }
573      if (!x->skip_encode && *eob)
574        vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, *eob);
575      break;
576    case TX_16X16:
577      tx_type = get_tx_type_16x16(pd->plane_type, xd);
578      scan = get_scan_16x16(tx_type);
579      iscan = get_iscan_16x16(tx_type);
580      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
581      block >>= 4;
582      xoff = 16 * (block & twmask);
583      yoff = 16 * (block >> twl);
584      dst = pd->dst.buf + yoff * pd->dst.stride + xoff;
585      vp9_predict_intra_block(xd, block, bwl, TX_16X16, mode,
586                              dst, pd->dst.stride, dst, pd->dst.stride);
587      if (!x->skip_recode) {
588        src = p->src.buf + yoff * p->src.stride + xoff;
589        src_diff = p->src_diff + 4 * bw * yoff + xoff;
590        vp9_subtract_block(16, 16, src_diff, bw * 4,
591                           src, p->src.stride, dst, pd->dst.stride);
592        vp9_fht16x16(tx_type, src_diff, coeff, bw * 4);
593        vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round,
594                       p->quant, p->quant_shift, qcoeff, dqcoeff,
595                       pd->dequant, p->zbin_extra, eob, scan, iscan);
596      }
597      if (!x->skip_encode && *eob)
598        vp9_iht16x16_add(tx_type, dqcoeff, dst, pd->dst.stride, *eob);
599      break;
600    case TX_8X8:
601      tx_type = get_tx_type_8x8(pd->plane_type, xd);
602      scan = get_scan_8x8(tx_type);
603      iscan = get_iscan_8x8(tx_type);
604      mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
605      block >>= 2;
606      xoff = 8 * (block & twmask);
607      yoff = 8 * (block >> twl);
608      dst = pd->dst.buf + yoff * pd->dst.stride + xoff;
609      vp9_predict_intra_block(xd, block, bwl, TX_8X8, mode,
610                              dst, pd->dst.stride, dst, pd->dst.stride);
611      if (!x->skip_recode) {
612        src = p->src.buf + yoff * p->src.stride + xoff;
613        src_diff = p->src_diff + 4 * bw * yoff + xoff;
614        vp9_subtract_block(8, 8, src_diff, bw * 4,
615                           src, p->src.stride, dst, pd->dst.stride);
616        vp9_fht8x8(tx_type, src_diff, coeff, bw * 4);
617        vp9_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round, p->quant,
618                       p->quant_shift, qcoeff, dqcoeff,
619                       pd->dequant, p->zbin_extra, eob, scan, iscan);
620      }
621      if (!x->skip_encode && *eob)
622        vp9_iht8x8_add(tx_type, dqcoeff, dst, pd->dst.stride, *eob);
623      break;
624    case TX_4X4:
625      tx_type = get_tx_type_4x4(pd->plane_type, xd, block);
626      scan = get_scan_4x4(tx_type);
627      iscan = get_iscan_4x4(tx_type);
628      if (mbmi->sb_type < BLOCK_8X8 && plane == 0)
629        mode = xd->mi_8x8[0]->bmi[block].as_mode;
630      else
631        mode = plane == 0 ? mbmi->mode : mbmi->uv_mode;
632
633      xoff = 4 * (block & twmask);
634      yoff = 4 * (block >> twl);
635      dst = pd->dst.buf + yoff * pd->dst.stride + xoff;
636      vp9_predict_intra_block(xd, block, bwl, TX_4X4, mode,
637                              dst, pd->dst.stride, dst, pd->dst.stride);
638
639      if (!x->skip_recode) {
640        src = p->src.buf + yoff * p->src.stride + xoff;
641        src_diff = p->src_diff + 4 * bw * yoff + xoff;
642        vp9_subtract_block(4, 4, src_diff, bw * 4,
643                           src, p->src.stride, dst, pd->dst.stride);
644        if (tx_type != DCT_DCT)
645          vp9_short_fht4x4(src_diff, coeff, bw * 4, tx_type);
646        else
647          x->fwd_txm4x4(src_diff, coeff, bw * 4);
648        vp9_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round, p->quant,
649                       p->quant_shift, qcoeff, dqcoeff,
650                       pd->dequant, p->zbin_extra, eob, scan, iscan);
651      }
652
653      if (!x->skip_encode && *eob) {
654        if (tx_type == DCT_DCT)
655          // this is like vp9_short_idct4x4 but has a special case around eob<=1
656          // which is significant (not just an optimization) for the lossless
657          // case.
658          xd->itxm_add(dqcoeff, dst, pd->dst.stride, *eob);
659        else
660          vp9_iht4x4_16_add(dqcoeff, dst, pd->dst.stride, tx_type);
661      }
662      break;
663    default:
664      assert(0);
665  }
666}
667
668void vp9_encode_intra_block_y(MACROBLOCK *x, BLOCK_SIZE bsize) {
669  MACROBLOCKD* const xd = &x->e_mbd;
670  struct optimize_ctx ctx;
671  struct encode_b_args arg = {x, &ctx};
672
673  foreach_transformed_block_in_plane(xd, bsize, 0, vp9_encode_block_intra,
674                                     &arg);
675}
676void vp9_encode_intra_block_uv(MACROBLOCK *x, BLOCK_SIZE bsize) {
677  MACROBLOCKD* const xd = &x->e_mbd;
678  struct optimize_ctx ctx;
679  struct encode_b_args arg = {x, &ctx};
680  foreach_transformed_block_uv(xd, bsize, vp9_encode_block_intra, &arg);
681}
682
683