2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
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.
14 #include "./vp9_rtcd.h"
16 #include "vpx_mem/vpx_mem.h"
18 #include "vp9/common/vp9_common.h"
19 #include "vp9/common/vp9_entropy.h"
20 #include "vp9/common/vp9_entropymode.h"
21 #include "vp9/common/vp9_idct.h"
22 #include "vp9/common/vp9_mvref_common.h"
23 #include "vp9/common/vp9_pred_common.h"
24 #include "vp9/common/vp9_quant_common.h"
25 #include "vp9/common/vp9_reconinter.h"
26 #include "vp9/common/vp9_reconintra.h"
27 #include "vp9/common/vp9_seg_common.h"
28 #include "vp9/common/vp9_systemdependent.h"
30 #include "vp9/encoder/vp9_cost.h"
31 #include "vp9/encoder/vp9_encodemb.h"
32 #include "vp9/encoder/vp9_encodemv.h"
33 #include "vp9/encoder/vp9_encoder.h"
34 #include "vp9/encoder/vp9_mcomp.h"
35 #include "vp9/encoder/vp9_quantize.h"
36 #include "vp9/encoder/vp9_ratectrl.h"
37 #include "vp9/encoder/vp9_rd.h"
38 #include "vp9/encoder/vp9_rdopt.h"
39 #include "vp9/encoder/vp9_variance.h"
41 #define RD_THRESH_MAX_FACT 64
42 #define RD_THRESH_INC 1
44 #define LAST_FRAME_MODE_MASK ((1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME) | \
46 #define GOLDEN_FRAME_MODE_MASK ((1 << LAST_FRAME) | (1 << ALTREF_FRAME) | \
48 #define ALT_REF_MODE_MASK ((1 << LAST_FRAME) | (1 << GOLDEN_FRAME) | \
51 #define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | 0x01)
53 #define MIN_EARLY_TERM_INDEX 3
57 MV_REFERENCE_FRAME ref_frame[2];
61 MV_REFERENCE_FRAME ref_frame[2];
64 struct rdcost_block_args {
66 ENTROPY_CONTEXT t_above[16];
67 ENTROPY_CONTEXT t_left[16];
77 int use_fast_coef_costing;
81 static const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
82 {NEARESTMV, {LAST_FRAME, NONE}},
83 {NEARESTMV, {ALTREF_FRAME, NONE}},
84 {NEARESTMV, {GOLDEN_FRAME, NONE}},
86 {DC_PRED, {INTRA_FRAME, NONE}},
88 {NEWMV, {LAST_FRAME, NONE}},
89 {NEWMV, {ALTREF_FRAME, NONE}},
90 {NEWMV, {GOLDEN_FRAME, NONE}},
92 {NEARMV, {LAST_FRAME, NONE}},
93 {NEARMV, {ALTREF_FRAME, NONE}},
94 {NEARESTMV, {LAST_FRAME, ALTREF_FRAME}},
95 {NEARESTMV, {GOLDEN_FRAME, ALTREF_FRAME}},
97 {TM_PRED, {INTRA_FRAME, NONE}},
99 {NEARMV, {LAST_FRAME, ALTREF_FRAME}},
100 {NEWMV, {LAST_FRAME, ALTREF_FRAME}},
101 {NEARMV, {GOLDEN_FRAME, NONE}},
102 {NEARMV, {GOLDEN_FRAME, ALTREF_FRAME}},
103 {NEWMV, {GOLDEN_FRAME, ALTREF_FRAME}},
105 {ZEROMV, {LAST_FRAME, NONE}},
106 {ZEROMV, {GOLDEN_FRAME, NONE}},
107 {ZEROMV, {ALTREF_FRAME, NONE}},
108 {ZEROMV, {LAST_FRAME, ALTREF_FRAME}},
109 {ZEROMV, {GOLDEN_FRAME, ALTREF_FRAME}},
111 {H_PRED, {INTRA_FRAME, NONE}},
112 {V_PRED, {INTRA_FRAME, NONE}},
113 {D135_PRED, {INTRA_FRAME, NONE}},
114 {D207_PRED, {INTRA_FRAME, NONE}},
115 {D153_PRED, {INTRA_FRAME, NONE}},
116 {D63_PRED, {INTRA_FRAME, NONE}},
117 {D117_PRED, {INTRA_FRAME, NONE}},
118 {D45_PRED, {INTRA_FRAME, NONE}},
121 static const REF_DEFINITION vp9_ref_order[MAX_REFS] = {
122 {{LAST_FRAME, NONE}},
123 {{GOLDEN_FRAME, NONE}},
124 {{ALTREF_FRAME, NONE}},
125 {{LAST_FRAME, ALTREF_FRAME}},
126 {{GOLDEN_FRAME, ALTREF_FRAME}},
127 {{INTRA_FRAME, NONE}},
130 static int raster_block_offset(BLOCK_SIZE plane_bsize,
131 int raster_block, int stride) {
132 const int bw = b_width_log2(plane_bsize);
133 const int y = 4 * (raster_block >> bw);
134 const int x = 4 * (raster_block & ((1 << bw) - 1));
135 return y * stride + x;
137 static int16_t* raster_block_offset_int16(BLOCK_SIZE plane_bsize,
138 int raster_block, int16_t *base) {
139 const int stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
140 return base + raster_block_offset(plane_bsize, raster_block, stride);
143 static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
144 int m, int n, int min_plane, int max_plane) {
147 for (i = min_plane; i < max_plane; ++i) {
148 struct macroblock_plane *const p = &x->plane[i];
149 struct macroblockd_plane *const pd = &x->e_mbd.plane[i];
151 p->coeff = ctx->coeff_pbuf[i][m];
152 p->qcoeff = ctx->qcoeff_pbuf[i][m];
153 pd->dqcoeff = ctx->dqcoeff_pbuf[i][m];
154 p->eobs = ctx->eobs_pbuf[i][m];
156 ctx->coeff_pbuf[i][m] = ctx->coeff_pbuf[i][n];
157 ctx->qcoeff_pbuf[i][m] = ctx->qcoeff_pbuf[i][n];
158 ctx->dqcoeff_pbuf[i][m] = ctx->dqcoeff_pbuf[i][n];
159 ctx->eobs_pbuf[i][m] = ctx->eobs_pbuf[i][n];
161 ctx->coeff_pbuf[i][n] = p->coeff;
162 ctx->qcoeff_pbuf[i][n] = p->qcoeff;
163 ctx->dqcoeff_pbuf[i][n] = pd->dqcoeff;
164 ctx->eobs_pbuf[i][n] = p->eobs;
168 static void model_rd_for_sb(VP9_COMP *cpi, BLOCK_SIZE bsize,
169 MACROBLOCK *x, MACROBLOCKD *xd,
170 int *out_rate_sum, int64_t *out_dist_sum) {
171 // Note our transform coeffs are 8 times an orthogonal transform.
172 // Hence quantizer step is also 8 times. To get effective quantizer
173 // we need to divide by 8 before sending to modeling function.
175 int64_t rate_sum = 0;
176 int64_t dist_sum = 0;
177 const int ref = xd->mi[0]->mbmi.ref_frame[0];
179 unsigned int var = 0;
180 unsigned int sum_sse = 0;
185 x->pred_sse[ref] = 0;
187 for (i = 0; i < MAX_MB_PLANE; ++i) {
188 struct macroblock_plane *const p = &x->plane[i];
189 struct macroblockd_plane *const pd = &xd->plane[i];
190 const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
191 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
192 const BLOCK_SIZE unit_size = txsize_to_bsize[max_tx_size];
193 int bw = 1 << (b_width_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
194 int bh = 1 << (b_height_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
196 int lw = b_width_log2_lookup[unit_size] + 2;
197 int lh = b_height_log2_lookup[unit_size] + 2;
201 for (idy = 0; idy < bh; ++idy) {
202 for (idx = 0; idx < bw; ++idx) {
203 uint8_t *src = p->src.buf + (idy * p->src.stride << lh) + (idx << lw);
204 uint8_t *dst = pd->dst.buf + (idy * pd->dst.stride << lh) + (idx << lh);
205 int block_idx = (idy << 1) + idx;
207 var = cpi->fn_ptr[unit_size].vf(src, p->src.stride,
208 dst, pd->dst.stride, &sse);
209 x->bsse[(i << 2) + block_idx] = sse;
212 if (!x->select_tx_size) {
213 if (x->bsse[(i << 2) + block_idx] < p->quant_thred[0] >> shift)
214 x->skip_txfm[(i << 2) + block_idx] = 1;
215 else if (var < p->quant_thred[1] >> shift)
216 x->skip_txfm[(i << 2) + block_idx] = 2;
218 x->skip_txfm[(i << 2) + block_idx] = 0;
222 x->pred_sse[ref] += sse;
226 // Fast approximate the modelling function.
227 if (cpi->oxcf.speed > 4) {
230 int64_t square_error = sse;
231 int quantizer = (pd->dequant[1] >> 3);
234 rate = (square_error * (280 - quantizer)) >> 8;
237 dist = (square_error * quantizer) >> 8;
241 vp9_model_rd_from_var_lapndz(sum_sse, 1 << num_pels_log2_lookup[bs],
242 pd->dequant[1] >> 3, &rate, &dist);
248 *out_rate_sum = (int)rate_sum;
249 *out_dist_sum = dist_sum << 4;
252 int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
253 intptr_t block_size, int64_t *ssz) {
255 int64_t error = 0, sqcoeff = 0;
257 for (i = 0; i < block_size; i++) {
258 const int diff = coeff[i] - dqcoeff[i];
259 error += diff * diff;
260 sqcoeff += coeff[i] * coeff[i];
267 /* The trailing '0' is a terminator which is used inside cost_coeffs() to
268 * decide whether to include cost of a trailing EOB node or not (i.e. we
269 * can skip this if the last coefficient in this transform block, e.g. the
270 * 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block,
272 static const int16_t band_counts[TX_SIZES][8] = {
273 { 1, 2, 3, 4, 3, 16 - 13, 0 },
274 { 1, 2, 3, 4, 11, 64 - 21, 0 },
275 { 1, 2, 3, 4, 11, 256 - 21, 0 },
276 { 1, 2, 3, 4, 11, 1024 - 21, 0 },
278 static INLINE int cost_coeffs(MACROBLOCK *x,
279 int plane, int block,
280 ENTROPY_CONTEXT *A, ENTROPY_CONTEXT *L,
282 const int16_t *scan, const int16_t *nb,
283 int use_fast_coef_costing) {
284 MACROBLOCKD *const xd = &x->e_mbd;
285 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
286 const struct macroblock_plane *p = &x->plane[plane];
287 const struct macroblockd_plane *pd = &xd->plane[plane];
288 const PLANE_TYPE type = pd->plane_type;
289 const int16_t *band_count = &band_counts[tx_size][1];
290 const int eob = p->eobs[block];
291 const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
292 unsigned int (*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
293 x->token_costs[tx_size][type][is_inter_block(mbmi)];
294 uint8_t token_cache[32 * 32];
295 int pt = combine_entropy_contexts(*A, *L);
297 // Check for consistency of tx_size with mode info
298 assert(type == PLANE_TYPE_Y ? mbmi->tx_size == tx_size
299 : get_uv_tx_size(mbmi, pd) == tx_size);
303 cost = token_costs[0][0][pt][EOB_TOKEN];
306 int band_left = *band_count++;
310 int prev_t = vp9_dct_value_tokens_ptr[v].token;
311 cost = (*token_costs)[0][pt][prev_t] + vp9_dct_value_cost_ptr[v];
312 token_cache[0] = vp9_pt_energy_class[prev_t];
316 for (c = 1; c < eob; c++) {
317 const int rc = scan[c];
321 t = vp9_dct_value_tokens_ptr[v].token;
322 if (use_fast_coef_costing) {
323 cost += (*token_costs)[!prev_t][!prev_t][t] + vp9_dct_value_cost_ptr[v];
325 pt = get_coef_context(nb, token_cache, c);
326 cost += (*token_costs)[!prev_t][pt][t] + vp9_dct_value_cost_ptr[v];
327 token_cache[rc] = vp9_pt_energy_class[t];
331 band_left = *band_count++;
338 if (use_fast_coef_costing) {
339 cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
341 pt = get_coef_context(nb, token_cache, c);
342 cost += (*token_costs)[0][pt][EOB_TOKEN];
347 // is eob first coefficient;
352 static void dist_block(int plane, int block, TX_SIZE tx_size,
353 struct rdcost_block_args* args) {
354 const int ss_txfrm_size = tx_size << 1;
355 MACROBLOCK* const x = args->x;
356 MACROBLOCKD* const xd = &x->e_mbd;
357 const struct macroblock_plane *const p = &x->plane[plane];
358 const struct macroblockd_plane *const pd = &xd->plane[plane];
360 int shift = tx_size == TX_32X32 ? 0 : 2;
361 tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
362 tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
363 args->dist = vp9_block_error(coeff, dqcoeff, 16 << ss_txfrm_size,
365 args->sse = this_sse >> shift;
367 if (x->skip_encode && !is_inter_block(&xd->mi[0]->mbmi)) {
368 // TODO(jingning): tune the model to better capture the distortion.
369 int64_t p = (pd->dequant[1] * pd->dequant[1] *
370 (1 << ss_txfrm_size)) >> (shift + 2);
371 args->dist += (p >> 4);
376 static void rate_block(int plane, int block, BLOCK_SIZE plane_bsize,
377 TX_SIZE tx_size, struct rdcost_block_args* args) {
379 txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x_idx, &y_idx);
381 args->rate = cost_coeffs(args->x, plane, block, args->t_above + x_idx,
382 args->t_left + y_idx, tx_size,
383 args->so->scan, args->so->neighbors,
384 args->use_fast_coef_costing);
387 static void block_rd_txfm(int plane, int block, BLOCK_SIZE plane_bsize,
388 TX_SIZE tx_size, void *arg) {
389 struct rdcost_block_args *args = arg;
390 MACROBLOCK *const x = args->x;
391 MACROBLOCKD *const xd = &x->e_mbd;
392 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
393 int64_t rd1, rd2, rd;
398 if (!is_inter_block(mbmi)) {
399 vp9_encode_block_intra(x, plane, block, plane_bsize, tx_size, &mbmi->skip);
400 dist_block(plane, block, tx_size, args);
401 } else if (max_txsize_lookup[plane_bsize] == tx_size) {
402 if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] == 0) {
403 // full forward transform and quantization
404 vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
405 dist_block(plane, block, tx_size, args);
406 } else if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] == 2) {
407 // compute DC coefficient
408 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[plane].coeff, block);
409 tran_low_t *const dqcoeff = BLOCK_OFFSET(xd->plane[plane].dqcoeff, block);
410 vp9_xform_quant_dc(x, plane, block, plane_bsize, tx_size);
411 args->sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
412 args->dist = args->sse;
413 if (!x->plane[plane].eobs[block])
414 args->dist = args->sse - ((coeff[0] * coeff[0] -
415 (coeff[0] - dqcoeff[0]) * (coeff[0] - dqcoeff[0])) >> 2);
417 // skip forward transform
418 x->plane[plane].eobs[block] = 0;
419 args->sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
420 args->dist = args->sse;
423 // full forward transform and quantization
424 vp9_xform_quant(x, plane, block, plane_bsize, tx_size);
425 dist_block(plane, block, tx_size, args);
428 rate_block(plane, block, plane_bsize, tx_size, args);
429 rd1 = RDCOST(x->rdmult, x->rddiv, args->rate, args->dist);
430 rd2 = RDCOST(x->rdmult, x->rddiv, 0, args->sse);
432 // TODO(jingning): temporarily enabled only for luma component
435 x->zcoeff_blk[tx_size][block] = !x->plane[plane].eobs[block] ||
436 (rd1 > rd2 && !xd->lossless);
438 args->this_rate += args->rate;
439 args->this_dist += args->dist;
440 args->this_sse += args->sse;
443 if (args->this_rd > args->best_rd) {
449 static void txfm_rd_in_plane(MACROBLOCK *x,
450 int *rate, int64_t *distortion,
451 int *skippable, int64_t *sse,
452 int64_t ref_best_rd, int plane,
453 BLOCK_SIZE bsize, TX_SIZE tx_size,
454 int use_fast_coef_casting) {
455 MACROBLOCKD *const xd = &x->e_mbd;
456 const struct macroblockd_plane *const pd = &xd->plane[plane];
457 struct rdcost_block_args args;
460 args.best_rd = ref_best_rd;
461 args.use_fast_coef_costing = use_fast_coef_casting;
464 xd->mi[0]->mbmi.tx_size = tx_size;
466 vp9_get_entropy_contexts(bsize, tx_size, pd, args.t_above, args.t_left);
468 args.so = get_scan(xd, tx_size, pd->plane_type, 0);
470 vp9_foreach_transformed_block_in_plane(xd, bsize, plane,
471 block_rd_txfm, &args);
474 *distortion = INT64_MAX;
478 *distortion = args.this_dist;
479 *rate = args.this_rate;
480 *sse = args.this_sse;
481 *skippable = vp9_is_skippable_in_plane(x, bsize, plane);
485 static void choose_largest_tx_size(VP9_COMP *cpi, MACROBLOCK *x,
486 int *rate, int64_t *distortion,
487 int *skip, int64_t *sse,
490 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
491 VP9_COMMON *const cm = &cpi->common;
492 const TX_SIZE largest_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
493 MACROBLOCKD *const xd = &x->e_mbd;
494 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
496 mbmi->tx_size = MIN(max_tx_size, largest_tx_size);
498 txfm_rd_in_plane(x, rate, distortion, skip,
499 sse, ref_best_rd, 0, bs,
500 mbmi->tx_size, cpi->sf.use_fast_coef_costing);
503 static void choose_tx_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x,
508 int64_t tx_cache[TX_MODES],
511 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
512 VP9_COMMON *const cm = &cpi->common;
513 MACROBLOCKD *const xd = &x->e_mbd;
514 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
515 vp9_prob skip_prob = vp9_get_skip_prob(cm, xd);
516 int r[TX_SIZES][2], s[TX_SIZES];
517 int64_t d[TX_SIZES], sse[TX_SIZES];
518 int64_t rd[TX_SIZES][2] = {{INT64_MAX, INT64_MAX},
519 {INT64_MAX, INT64_MAX},
520 {INT64_MAX, INT64_MAX},
521 {INT64_MAX, INT64_MAX}};
524 const TX_SIZE max_mode_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
525 int64_t best_rd = INT64_MAX;
526 TX_SIZE best_tx = max_tx_size;
528 const vp9_prob *tx_probs = get_tx_probs2(max_tx_size, xd, &cm->fc.tx_probs);
529 assert(skip_prob > 0);
530 s0 = vp9_cost_bit(skip_prob, 0);
531 s1 = vp9_cost_bit(skip_prob, 1);
533 for (n = max_tx_size; n >= 0; n--) {
534 txfm_rd_in_plane(x, &r[n][0], &d[n], &s[n],
535 &sse[n], ref_best_rd, 0, bs, n,
536 cpi->sf.use_fast_coef_costing);
538 if (r[n][0] < INT_MAX) {
539 for (m = 0; m <= n - (n == (int) max_tx_size); m++) {
541 r[n][1] += vp9_cost_zero(tx_probs[m]);
543 r[n][1] += vp9_cost_one(tx_probs[m]);
546 if (d[n] == INT64_MAX) {
547 rd[n][0] = rd[n][1] = INT64_MAX;
549 rd[n][0] = rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1, d[n]);
551 rd[n][0] = RDCOST(x->rdmult, x->rddiv, r[n][0] + s0, d[n]);
552 rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]);
555 // Early termination in transform size search.
556 if (cpi->sf.tx_size_search_breakout &&
557 (rd[n][1] == INT64_MAX ||
558 (n < (int) max_tx_size && rd[n][1] > rd[n + 1][1]) ||
562 if (rd[n][1] < best_rd) {
567 mbmi->tx_size = cm->tx_mode == TX_MODE_SELECT ?
568 best_tx : MIN(max_tx_size, max_mode_tx_size);
571 *distortion = d[mbmi->tx_size];
572 *rate = r[mbmi->tx_size][cm->tx_mode == TX_MODE_SELECT];
573 *skip = s[mbmi->tx_size];
574 *psse = sse[mbmi->tx_size];
576 tx_cache[ONLY_4X4] = rd[TX_4X4][0];
577 tx_cache[ALLOW_8X8] = rd[TX_8X8][0];
578 tx_cache[ALLOW_16X16] = rd[MIN(max_tx_size, TX_16X16)][0];
579 tx_cache[ALLOW_32X32] = rd[MIN(max_tx_size, TX_32X32)][0];
581 if (max_tx_size == TX_32X32 && best_tx == TX_32X32) {
582 tx_cache[TX_MODE_SELECT] = rd[TX_32X32][1];
583 } else if (max_tx_size >= TX_16X16 && best_tx == TX_16X16) {
584 tx_cache[TX_MODE_SELECT] = rd[TX_16X16][1];
585 } else if (rd[TX_8X8][1] < rd[TX_4X4][1]) {
586 tx_cache[TX_MODE_SELECT] = rd[TX_8X8][1];
588 tx_cache[TX_MODE_SELECT] = rd[TX_4X4][1];
592 static void super_block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
593 int64_t *distortion, int *skip,
594 int64_t *psse, BLOCK_SIZE bs,
595 int64_t txfm_cache[TX_MODES],
596 int64_t ref_best_rd) {
597 MACROBLOCKD *xd = &x->e_mbd;
599 int64_t *ret_sse = psse ? psse : &sse;
601 assert(bs == xd->mi[0]->mbmi.sb_type);
603 if (cpi->sf.tx_size_search_method == USE_LARGESTALL || xd->lossless) {
604 vpx_memset(txfm_cache, 0, TX_MODES * sizeof(int64_t));
605 choose_largest_tx_size(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
608 choose_tx_size_from_rd(cpi, x, rate, distortion, skip, ret_sse,
609 txfm_cache, ref_best_rd, bs);
613 static int conditional_skipintra(PREDICTION_MODE mode,
614 PREDICTION_MODE best_intra_mode) {
615 if (mode == D117_PRED &&
616 best_intra_mode != V_PRED &&
617 best_intra_mode != D135_PRED)
619 if (mode == D63_PRED &&
620 best_intra_mode != V_PRED &&
621 best_intra_mode != D45_PRED)
623 if (mode == D207_PRED &&
624 best_intra_mode != H_PRED &&
625 best_intra_mode != D45_PRED)
627 if (mode == D153_PRED &&
628 best_intra_mode != H_PRED &&
629 best_intra_mode != D135_PRED)
634 static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x, int ib,
635 PREDICTION_MODE *best_mode,
636 const int *bmode_costs,
637 ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l,
638 int *bestrate, int *bestratey,
639 int64_t *bestdistortion,
640 BLOCK_SIZE bsize, int64_t rd_thresh) {
641 PREDICTION_MODE mode;
642 MACROBLOCKD *const xd = &x->e_mbd;
643 int64_t best_rd = rd_thresh;
645 struct macroblock_plane *p = &x->plane[0];
646 struct macroblockd_plane *pd = &xd->plane[0];
647 const int src_stride = p->src.stride;
648 const int dst_stride = pd->dst.stride;
649 const uint8_t *src_init = &p->src.buf[raster_block_offset(BLOCK_8X8, ib,
651 uint8_t *dst_init = &pd->dst.buf[raster_block_offset(BLOCK_8X8, ib,
653 ENTROPY_CONTEXT ta[2], tempa[2];
654 ENTROPY_CONTEXT tl[2], templ[2];
656 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
657 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
659 uint8_t best_dst[8 * 8];
663 vpx_memcpy(ta, a, sizeof(ta));
664 vpx_memcpy(tl, l, sizeof(tl));
665 xd->mi[0]->mbmi.tx_size = TX_4X4;
667 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
670 int64_t distortion = 0;
671 int rate = bmode_costs[mode];
673 if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode)))
676 // Only do the oblique modes if the best so far is
677 // one of the neighboring directional modes
678 if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
679 if (conditional_skipintra(mode, *best_mode))
683 vpx_memcpy(tempa, ta, sizeof(ta));
684 vpx_memcpy(templ, tl, sizeof(tl));
686 for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
687 for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
688 const int block = ib + idy * 2 + idx;
689 const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
690 uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
691 int16_t *const src_diff = raster_block_offset_int16(BLOCK_8X8, block,
693 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
694 xd->mi[0]->bmi[block].as_mode = mode;
695 vp9_predict_intra_block(xd, block, 1,
697 x->skip_encode ? src : dst,
698 x->skip_encode ? src_stride : dst_stride,
699 dst, dst_stride, idx, idy, 0);
700 vp9_subtract_block(4, 4, src_diff, 8, src, src_stride, dst, dst_stride);
703 const scan_order *so = &vp9_default_scan_orders[TX_4X4];
704 vp9_fwht4x4(src_diff, coeff, 8);
705 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
706 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
707 so->scan, so->neighbors,
708 cpi->sf.use_fast_coef_costing);
709 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
711 vp9_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block), dst, dst_stride,
715 const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
716 const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
717 vp9_fht4x4(src_diff, coeff, 8, tx_type);
718 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
719 ratey += cost_coeffs(x, 0, block, tempa + idx, templ + idy, TX_4X4,
720 so->scan, so->neighbors,
721 cpi->sf.use_fast_coef_costing);
722 distortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, block),
724 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
726 vp9_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block),
727 dst, dst_stride, p->eobs[block]);
733 this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
735 if (this_rd < best_rd) {
738 *bestdistortion = distortion;
741 vpx_memcpy(a, tempa, sizeof(tempa));
742 vpx_memcpy(l, templ, sizeof(templ));
743 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
744 vpx_memcpy(best_dst + idy * 8, dst_init + idy * dst_stride,
745 num_4x4_blocks_wide * 4);
751 if (best_rd >= rd_thresh || x->skip_encode)
754 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
755 vpx_memcpy(dst_init + idy * dst_stride, best_dst + idy * 8,
756 num_4x4_blocks_wide * 4);
761 static int64_t rd_pick_intra_sub_8x8_y_mode(VP9_COMP *cpi, MACROBLOCK *mb,
762 int *rate, int *rate_y,
766 const MACROBLOCKD *const xd = &mb->e_mbd;
767 MODE_INFO *const mic = xd->mi[0];
768 const MODE_INFO *above_mi = xd->mi[-xd->mi_stride];
769 const MODE_INFO *left_mi = xd->left_available ? xd->mi[-1] : NULL;
770 const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
771 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
772 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
775 int64_t total_distortion = 0;
777 int64_t total_rd = 0;
778 ENTROPY_CONTEXT t_above[4], t_left[4];
779 const int *bmode_costs = cpi->mbmode_cost;
781 vpx_memcpy(t_above, xd->plane[0].above_context, sizeof(t_above));
782 vpx_memcpy(t_left, xd->plane[0].left_context, sizeof(t_left));
784 // Pick modes for each sub-block (of size 4x4, 4x8, or 8x4) in an 8x8 block.
785 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
786 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
787 PREDICTION_MODE best_mode = DC_PRED;
788 int r = INT_MAX, ry = INT_MAX;
789 int64_t d = INT64_MAX, this_rd = INT64_MAX;
791 if (cpi->common.frame_type == KEY_FRAME) {
792 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, i);
793 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, i);
795 bmode_costs = cpi->y_mode_costs[A][L];
798 this_rd = rd_pick_intra4x4block(cpi, mb, i, &best_mode, bmode_costs,
799 t_above + idx, t_left + idy, &r, &ry, &d,
800 bsize, best_rd - total_rd);
801 if (this_rd >= best_rd - total_rd)
806 total_distortion += d;
809 mic->bmi[i].as_mode = best_mode;
810 for (j = 1; j < num_4x4_blocks_high; ++j)
811 mic->bmi[i + j * 2].as_mode = best_mode;
812 for (j = 1; j < num_4x4_blocks_wide; ++j)
813 mic->bmi[i + j].as_mode = best_mode;
815 if (total_rd >= best_rd)
821 *rate_y = tot_rate_y;
822 *distortion = total_distortion;
823 mic->mbmi.mode = mic->bmi[3].as_mode;
825 return RDCOST(mb->rdmult, mb->rddiv, cost, total_distortion);
828 static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x,
829 int *rate, int *rate_tokenonly,
830 int64_t *distortion, int *skippable,
832 int64_t tx_cache[TX_MODES],
834 PREDICTION_MODE mode;
835 PREDICTION_MODE mode_selected = DC_PRED;
836 MACROBLOCKD *const xd = &x->e_mbd;
837 MODE_INFO *const mic = xd->mi[0];
838 int this_rate, this_rate_tokenonly, s;
839 int64_t this_distortion, this_rd;
840 TX_SIZE best_tx = TX_4X4;
842 int *bmode_costs = cpi->mbmode_cost;
844 if (cpi->sf.tx_size_search_method == USE_FULL_RD)
845 for (i = 0; i < TX_MODES; i++)
846 tx_cache[i] = INT64_MAX;
848 /* Y Search for intra prediction mode */
849 for (mode = DC_PRED; mode <= TM_PRED; mode++) {
850 int64_t local_tx_cache[TX_MODES];
851 MODE_INFO *above_mi = xd->mi[-xd->mi_stride];
852 MODE_INFO *left_mi = xd->left_available ? xd->mi[-1] : NULL;
854 if (cpi->common.frame_type == KEY_FRAME) {
855 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
856 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
858 bmode_costs = cpi->y_mode_costs[A][L];
860 mic->mbmi.mode = mode;
862 super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion,
863 &s, NULL, bsize, local_tx_cache, best_rd);
865 if (this_rate_tokenonly == INT_MAX)
868 this_rate = this_rate_tokenonly + bmode_costs[mode];
869 this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
871 if (this_rd < best_rd) {
872 mode_selected = mode;
874 best_tx = mic->mbmi.tx_size;
876 *rate_tokenonly = this_rate_tokenonly;
877 *distortion = this_distortion;
881 if (cpi->sf.tx_size_search_method == USE_FULL_RD && this_rd < INT64_MAX) {
882 for (i = 0; i < TX_MODES && local_tx_cache[i] < INT64_MAX; i++) {
883 const int64_t adj_rd = this_rd + local_tx_cache[i] -
884 local_tx_cache[cpi->common.tx_mode];
885 if (adj_rd < tx_cache[i]) {
886 tx_cache[i] = adj_rd;
892 mic->mbmi.mode = mode_selected;
893 mic->mbmi.tx_size = best_tx;
898 static void super_block_uvrd(const VP9_COMP *cpi, MACROBLOCK *x,
899 int *rate, int64_t *distortion, int *skippable,
900 int64_t *sse, BLOCK_SIZE bsize,
901 int64_t ref_best_rd) {
902 MACROBLOCKD *const xd = &x->e_mbd;
903 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
904 const TX_SIZE uv_tx_size = get_uv_tx_size(mbmi, &xd->plane[1]);
906 int pnrate = 0, pnskip = 1;
907 int64_t pndist = 0, pnsse = 0;
912 if (is_inter_block(mbmi)) {
914 for (plane = 1; plane < MAX_MB_PLANE; ++plane)
915 vp9_subtract_plane(x, bsize, plane);
923 for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
924 txfm_rd_in_plane(x, &pnrate, &pndist, &pnskip, &pnsse,
925 ref_best_rd, plane, bsize, uv_tx_size,
926 cpi->sf.use_fast_coef_costing);
927 if (pnrate == INT_MAX)
930 *distortion += pndist;
932 *skippable &= pnskip;
938 *distortion = INT64_MAX;
944 static int64_t rd_pick_intra_sbuv_mode(VP9_COMP *cpi, MACROBLOCK *x,
945 PICK_MODE_CONTEXT *ctx,
946 int *rate, int *rate_tokenonly,
947 int64_t *distortion, int *skippable,
948 BLOCK_SIZE bsize, TX_SIZE max_tx_size) {
949 MACROBLOCKD *xd = &x->e_mbd;
950 PREDICTION_MODE mode;
951 PREDICTION_MODE mode_selected = DC_PRED;
952 int64_t best_rd = INT64_MAX, this_rd;
953 int this_rate_tokenonly, this_rate, s;
954 int64_t this_distortion, this_sse;
956 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
957 if (!(cpi->sf.intra_uv_mode_mask[max_tx_size] & (1 << mode)))
960 xd->mi[0]->mbmi.uv_mode = mode;
962 super_block_uvrd(cpi, x, &this_rate_tokenonly,
963 &this_distortion, &s, &this_sse, bsize, best_rd);
964 if (this_rate_tokenonly == INT_MAX)
966 this_rate = this_rate_tokenonly +
967 cpi->intra_uv_mode_cost[cpi->common.frame_type][mode];
968 this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
970 if (this_rd < best_rd) {
971 mode_selected = mode;
974 *rate_tokenonly = this_rate_tokenonly;
975 *distortion = this_distortion;
977 if (!x->select_tx_size)
978 swap_block_ptr(x, ctx, 2, 0, 1, MAX_MB_PLANE);
982 xd->mi[0]->mbmi.uv_mode = mode_selected;
986 static int64_t rd_sbuv_dcpred(const VP9_COMP *cpi, MACROBLOCK *x,
987 int *rate, int *rate_tokenonly,
988 int64_t *distortion, int *skippable,
990 const VP9_COMMON *cm = &cpi->common;
993 x->e_mbd.mi[0]->mbmi.uv_mode = DC_PRED;
994 super_block_uvrd(cpi, x, rate_tokenonly, distortion,
995 skippable, &unused, bsize, INT64_MAX);
996 *rate = *rate_tokenonly + cpi->intra_uv_mode_cost[cm->frame_type][DC_PRED];
997 return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
1000 static void choose_intra_uv_mode(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx,
1001 BLOCK_SIZE bsize, TX_SIZE max_tx_size,
1002 int *rate_uv, int *rate_uv_tokenonly,
1003 int64_t *dist_uv, int *skip_uv,
1004 PREDICTION_MODE *mode_uv) {
1005 MACROBLOCK *const x = &cpi->mb;
1007 // Use an estimated rd for uv_intra based on DC_PRED if the
1008 // appropriate speed flag is set.
1009 if (cpi->sf.use_uv_intra_rd_estimate) {
1010 rd_sbuv_dcpred(cpi, x, rate_uv, rate_uv_tokenonly, dist_uv,
1011 skip_uv, bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize);
1012 // Else do a proper rd search for each possible transform size that may
1013 // be considered in the main rd loop.
1015 rd_pick_intra_sbuv_mode(cpi, x, ctx,
1016 rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
1017 bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize, max_tx_size);
1019 *mode_uv = x->e_mbd.mi[0]->mbmi.uv_mode;
1022 static int cost_mv_ref(const VP9_COMP *cpi, PREDICTION_MODE mode,
1024 assert(is_inter_mode(mode));
1025 return cpi->inter_mode_cost[mode_context][INTER_OFFSET(mode)];
1028 static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
1031 int mi_row, int mi_col,
1032 int_mv single_newmv[MAX_REF_FRAMES],
1035 static int set_and_cost_bmi_mvs(VP9_COMP *cpi, MACROBLOCKD *xd, int i,
1036 PREDICTION_MODE mode, int_mv this_mv[2],
1037 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1038 int_mv seg_mvs[MAX_REF_FRAMES],
1039 int_mv *best_ref_mv[2], const int *mvjcost,
1041 MODE_INFO *const mic = xd->mi[0];
1042 const MB_MODE_INFO *const mbmi = &mic->mbmi;
1045 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[mbmi->sb_type];
1046 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mbmi->sb_type];
1047 const int is_compound = has_second_ref(mbmi);
1051 this_mv[0].as_int = seg_mvs[mbmi->ref_frame[0]].as_int;
1052 thismvcost += vp9_mv_bit_cost(&this_mv[0].as_mv, &best_ref_mv[0]->as_mv,
1053 mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1055 this_mv[1].as_int = seg_mvs[mbmi->ref_frame[1]].as_int;
1056 thismvcost += vp9_mv_bit_cost(&this_mv[1].as_mv, &best_ref_mv[1]->as_mv,
1057 mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1062 this_mv[0].as_int = frame_mv[mode][mbmi->ref_frame[0]].as_int;
1064 this_mv[1].as_int = frame_mv[mode][mbmi->ref_frame[1]].as_int;
1067 this_mv[0].as_int = 0;
1069 this_mv[1].as_int = 0;
1075 mic->bmi[i].as_mv[0].as_int = this_mv[0].as_int;
1077 mic->bmi[i].as_mv[1].as_int = this_mv[1].as_int;
1079 mic->bmi[i].as_mode = mode;
1081 for (idy = 0; idy < num_4x4_blocks_high; ++idy)
1082 for (idx = 0; idx < num_4x4_blocks_wide; ++idx)
1083 vpx_memcpy(&mic->bmi[i + idy * 2 + idx],
1084 &mic->bmi[i], sizeof(mic->bmi[i]));
1086 return cost_mv_ref(cpi, mode, mbmi->mode_context[mbmi->ref_frame[0]]) +
1090 static int64_t encode_inter_mb_segment(VP9_COMP *cpi,
1095 int64_t *distortion, int64_t *sse,
1096 ENTROPY_CONTEXT *ta,
1097 ENTROPY_CONTEXT *tl,
1098 int mi_row, int mi_col) {
1100 MACROBLOCKD *xd = &x->e_mbd;
1101 struct macroblockd_plane *const pd = &xd->plane[0];
1102 struct macroblock_plane *const p = &x->plane[0];
1103 MODE_INFO *const mi = xd->mi[0];
1104 const BLOCK_SIZE plane_bsize = get_plane_block_size(mi->mbmi.sb_type, pd);
1105 const int width = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
1106 const int height = 4 * num_4x4_blocks_high_lookup[plane_bsize];
1109 const uint8_t *const src = &p->src.buf[raster_block_offset(BLOCK_8X8, i,
1111 uint8_t *const dst = &pd->dst.buf[raster_block_offset(BLOCK_8X8, i,
1113 int64_t thisdistortion = 0, thissse = 0;
1114 int thisrate = 0, ref;
1115 const scan_order *so = &vp9_default_scan_orders[TX_4X4];
1116 const int is_compound = has_second_ref(&mi->mbmi);
1117 const InterpKernel *kernel = vp9_get_interp_kernel(mi->mbmi.interp_filter);
1119 for (ref = 0; ref < 1 + is_compound; ++ref) {
1120 const uint8_t *pre = &pd->pre[ref].buf[raster_block_offset(BLOCK_8X8, i,
1121 pd->pre[ref].stride)];
1122 vp9_build_inter_predictor(pre, pd->pre[ref].stride,
1123 dst, pd->dst.stride,
1124 &mi->bmi[i].as_mv[ref].as_mv,
1125 &xd->block_refs[ref]->sf, width, height, ref,
1126 kernel, MV_PRECISION_Q3,
1127 mi_col * MI_SIZE + 4 * (i % 2),
1128 mi_row * MI_SIZE + 4 * (i / 2));
1131 vp9_subtract_block(height, width,
1132 raster_block_offset_int16(BLOCK_8X8, i, p->src_diff), 8,
1134 dst, pd->dst.stride);
1137 for (idy = 0; idy < height / 4; ++idy) {
1138 for (idx = 0; idx < width / 4; ++idx) {
1139 int64_t ssz, rd, rd1, rd2;
1142 k += (idy * 2 + idx);
1143 coeff = BLOCK_OFFSET(p->coeff, k);
1144 x->fwd_txm4x4(raster_block_offset_int16(BLOCK_8X8, k, p->src_diff),
1146 vp9_regular_quantize_b_4x4(x, 0, k, so->scan, so->iscan);
1147 thisdistortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k),
1150 thisrate += cost_coeffs(x, 0, k, ta + (k & 1), tl + (k >> 1), TX_4X4,
1151 so->scan, so->neighbors,
1152 cpi->sf.use_fast_coef_costing);
1153 rd1 = RDCOST(x->rdmult, x->rddiv, thisrate, thisdistortion >> 2);
1154 rd2 = RDCOST(x->rdmult, x->rddiv, 0, thissse >> 2);
1161 *distortion = thisdistortion >> 2;
1162 *labelyrate = thisrate;
1163 *sse = thissse >> 2;
1165 return RDCOST(x->rdmult, x->rddiv, *labelyrate, *distortion);
1176 ENTROPY_CONTEXT ta[2];
1177 ENTROPY_CONTEXT tl[2];
1189 PREDICTION_MODE modes[4];
1190 SEG_RDSTAT rdstat[4][INTER_MODES];
1194 static INLINE int mv_check_bounds(const MACROBLOCK *x, const MV *mv) {
1195 return (mv->row >> 3) < x->mv_row_min ||
1196 (mv->row >> 3) > x->mv_row_max ||
1197 (mv->col >> 3) < x->mv_col_min ||
1198 (mv->col >> 3) > x->mv_col_max;
1201 static INLINE void mi_buf_shift(MACROBLOCK *x, int i) {
1202 MB_MODE_INFO *const mbmi = &x->e_mbd.mi[0]->mbmi;
1203 struct macroblock_plane *const p = &x->plane[0];
1204 struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
1206 p->src.buf = &p->src.buf[raster_block_offset(BLOCK_8X8, i, p->src.stride)];
1207 assert(((intptr_t)pd->pre[0].buf & 0x7) == 0);
1208 pd->pre[0].buf = &pd->pre[0].buf[raster_block_offset(BLOCK_8X8, i,
1209 pd->pre[0].stride)];
1210 if (has_second_ref(mbmi))
1211 pd->pre[1].buf = &pd->pre[1].buf[raster_block_offset(BLOCK_8X8, i,
1212 pd->pre[1].stride)];
1215 static INLINE void mi_buf_restore(MACROBLOCK *x, struct buf_2d orig_src,
1216 struct buf_2d orig_pre[2]) {
1217 MB_MODE_INFO *mbmi = &x->e_mbd.mi[0]->mbmi;
1218 x->plane[0].src = orig_src;
1219 x->e_mbd.plane[0].pre[0] = orig_pre[0];
1220 if (has_second_ref(mbmi))
1221 x->e_mbd.plane[0].pre[1] = orig_pre[1];
1224 static INLINE int mv_has_subpel(const MV *mv) {
1225 return (mv->row & 0x0F) || (mv->col & 0x0F);
1228 // Check if NEARESTMV/NEARMV/ZEROMV is the cheapest way encode zero motion.
1229 // TODO(aconverse): Find out if this is still productive then clean up or remove
1230 static int check_best_zero_mv(
1231 const VP9_COMP *cpi, const uint8_t mode_context[MAX_REF_FRAMES],
1232 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], int this_mode,
1233 const MV_REFERENCE_FRAME ref_frames[2]) {
1234 if ((this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) &&
1235 frame_mv[this_mode][ref_frames[0]].as_int == 0 &&
1236 (ref_frames[1] == NONE ||
1237 frame_mv[this_mode][ref_frames[1]].as_int == 0)) {
1238 int rfc = mode_context[ref_frames[0]];
1239 int c1 = cost_mv_ref(cpi, NEARMV, rfc);
1240 int c2 = cost_mv_ref(cpi, NEARESTMV, rfc);
1241 int c3 = cost_mv_ref(cpi, ZEROMV, rfc);
1243 if (this_mode == NEARMV) {
1244 if (c1 > c3) return 0;
1245 } else if (this_mode == NEARESTMV) {
1246 if (c2 > c3) return 0;
1248 assert(this_mode == ZEROMV);
1249 if (ref_frames[1] == NONE) {
1250 if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0) ||
1251 (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0))
1254 if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0 &&
1255 frame_mv[NEARESTMV][ref_frames[1]].as_int == 0) ||
1256 (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0 &&
1257 frame_mv[NEARMV][ref_frames[1]].as_int == 0))
1265 static int64_t rd_pick_best_sub8x8_mode(VP9_COMP *cpi, MACROBLOCK *x,
1266 const TileInfo * const tile,
1267 int_mv *best_ref_mv,
1268 int_mv *second_best_ref_mv,
1269 int64_t best_rd, int *returntotrate,
1271 int64_t *returndistortion,
1272 int *skippable, int64_t *psse,
1274 int_mv seg_mvs[4][MAX_REF_FRAMES],
1275 BEST_SEG_INFO *bsi_buf, int filter_idx,
1276 int mi_row, int mi_col) {
1278 BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
1279 MACROBLOCKD *xd = &x->e_mbd;
1280 MODE_INFO *mi = xd->mi[0];
1281 MB_MODE_INFO *mbmi = &mi->mbmi;
1283 int k, br = 0, idx, idy;
1284 int64_t bd = 0, block_sse = 0;
1285 PREDICTION_MODE this_mode;
1286 VP9_COMMON *cm = &cpi->common;
1287 struct macroblock_plane *const p = &x->plane[0];
1288 struct macroblockd_plane *const pd = &xd->plane[0];
1289 const int label_count = 4;
1290 int64_t this_segment_rd = 0;
1291 int label_mv_thresh;
1292 int segmentyrate = 0;
1293 const BLOCK_SIZE bsize = mbmi->sb_type;
1294 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1295 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1296 ENTROPY_CONTEXT t_above[2], t_left[2];
1297 int subpelmv = 1, have_ref = 0;
1298 const int has_second_rf = has_second_ref(mbmi);
1299 const int inter_mode_mask = cpi->sf.inter_mode_mask[bsize];
1303 bsi->segment_rd = best_rd;
1304 bsi->ref_mv[0] = best_ref_mv;
1305 bsi->ref_mv[1] = second_best_ref_mv;
1306 bsi->mvp.as_int = best_ref_mv->as_int;
1307 bsi->mvthresh = mvthresh;
1309 for (i = 0; i < 4; i++)
1310 bsi->modes[i] = ZEROMV;
1312 vpx_memcpy(t_above, pd->above_context, sizeof(t_above));
1313 vpx_memcpy(t_left, pd->left_context, sizeof(t_left));
1315 // 64 makes this threshold really big effectively
1316 // making it so that we very rarely check mvs on
1317 // segments. setting this to 1 would make mv thresh
1318 // roughly equal to what it is for macroblocks
1319 label_mv_thresh = 1 * bsi->mvthresh / label_count;
1321 // Segmentation method overheads
1322 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1323 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1324 // TODO(jingning,rbultje): rewrite the rate-distortion optimization
1325 // loop for 4x4/4x8/8x4 block coding. to be replaced with new rd loop
1326 int_mv mode_mv[MB_MODE_COUNT][2];
1327 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1328 PREDICTION_MODE mode_selected = ZEROMV;
1329 int64_t best_rd = INT64_MAX;
1330 const int i = idy * 2 + idx;
1333 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1334 const MV_REFERENCE_FRAME frame = mbmi->ref_frame[ref];
1335 frame_mv[ZEROMV][frame].as_int = 0;
1336 vp9_append_sub8x8_mvs_for_idx(cm, xd, tile, i, ref, mi_row, mi_col,
1337 &frame_mv[NEARESTMV][frame],
1338 &frame_mv[NEARMV][frame]);
1341 // search for the best motion vector on this segment
1342 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
1343 const struct buf_2d orig_src = x->plane[0].src;
1344 struct buf_2d orig_pre[2];
1346 mode_idx = INTER_OFFSET(this_mode);
1347 bsi->rdstat[i][mode_idx].brdcost = INT64_MAX;
1348 if (!(inter_mode_mask & (1 << this_mode)))
1351 if (!check_best_zero_mv(cpi, mbmi->mode_context, frame_mv,
1352 this_mode, mbmi->ref_frame))
1355 vpx_memcpy(orig_pre, pd->pre, sizeof(orig_pre));
1356 vpx_memcpy(bsi->rdstat[i][mode_idx].ta, t_above,
1357 sizeof(bsi->rdstat[i][mode_idx].ta));
1358 vpx_memcpy(bsi->rdstat[i][mode_idx].tl, t_left,
1359 sizeof(bsi->rdstat[i][mode_idx].tl));
1361 // motion search for newmv (single predictor case only)
1362 if (!has_second_rf && this_mode == NEWMV &&
1363 seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV) {
1364 MV *const new_mv = &mode_mv[NEWMV][0].as_mv;
1366 int thissme, bestsme = INT_MAX;
1367 int sadpb = x->sadperbit4;
1372 /* Is the best so far sufficiently good that we cant justify doing
1373 * and new motion search. */
1374 if (best_rd < label_mv_thresh)
1377 if (cpi->oxcf.mode != BEST) {
1378 // use previous block's result as next block's MV predictor.
1380 bsi->mvp.as_int = mi->bmi[i - 1].as_mv[0].as_int;
1382 bsi->mvp.as_int = mi->bmi[i - 2].as_mv[0].as_int;
1386 max_mv = x->max_mv_context[mbmi->ref_frame[0]];
1388 max_mv = MAX(abs(bsi->mvp.as_mv.row), abs(bsi->mvp.as_mv.col)) >> 3;
1390 if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
1391 // Take wtd average of the step_params based on the last frame's
1392 // max mv magnitude and the best ref mvs of the current block for
1393 // the given reference.
1394 step_param = (vp9_init_search_range(max_mv) +
1395 cpi->mv_step_param) / 2;
1397 step_param = cpi->mv_step_param;
1400 mvp_full.row = bsi->mvp.as_mv.row >> 3;
1401 mvp_full.col = bsi->mvp.as_mv.col >> 3;
1403 if (cpi->sf.adaptive_motion_search) {
1404 mvp_full.row = x->pred_mv[mbmi->ref_frame[0]].row >> 3;
1405 mvp_full.col = x->pred_mv[mbmi->ref_frame[0]].col >> 3;
1406 step_param = MAX(step_param, 8);
1409 // adjust src pointer for this block
1412 vp9_set_mv_search_range(x, &bsi->ref_mv[0]->as_mv);
1414 bestsme = vp9_full_pixel_search(
1415 cpi, x, bsize, &mvp_full, step_param, sadpb,
1416 cpi->sf.mv.subpel_search_method != SUBPEL_TREE ? sad_list : NULL,
1417 &bsi->ref_mv[0]->as_mv, new_mv,
1420 // Should we do a full search (best quality only)
1421 if (cpi->oxcf.mode == BEST) {
1422 int_mv *const best_mv = &mi->bmi[i].as_mv[0];
1423 /* Check if mvp_full is within the range. */
1424 clamp_mv(&mvp_full, x->mv_col_min, x->mv_col_max,
1425 x->mv_row_min, x->mv_row_max);
1426 thissme = cpi->full_search_sad(x, &mvp_full,
1427 sadpb, 16, &cpi->fn_ptr[bsize],
1428 &bsi->ref_mv[0]->as_mv,
1430 sad_list[1] = sad_list[2] = sad_list[3] = sad_list[4] = INT_MAX;
1431 if (thissme < bestsme) {
1433 *new_mv = best_mv->as_mv;
1435 // The full search result is actually worse so re-instate the
1436 // previous best vector
1437 best_mv->as_mv = *new_mv;
1441 if (bestsme < INT_MAX) {
1443 cpi->find_fractional_mv_step(
1446 &bsi->ref_mv[0]->as_mv,
1447 cm->allow_high_precision_mv,
1448 x->errorperbit, &cpi->fn_ptr[bsize],
1449 cpi->sf.mv.subpel_force_stop,
1450 cpi->sf.mv.subpel_iters_per_step,
1451 cond_sad_list(cpi, sad_list),
1452 x->nmvjointcost, x->mvcost,
1454 &x->pred_sse[mbmi->ref_frame[0]],
1457 // save motion search result for use in compound prediction
1458 seg_mvs[i][mbmi->ref_frame[0]].as_mv = *new_mv;
1461 if (cpi->sf.adaptive_motion_search)
1462 x->pred_mv[mbmi->ref_frame[0]] = *new_mv;
1464 // restore src pointers
1465 mi_buf_restore(x, orig_src, orig_pre);
1468 if (has_second_rf) {
1469 if (seg_mvs[i][mbmi->ref_frame[1]].as_int == INVALID_MV ||
1470 seg_mvs[i][mbmi->ref_frame[0]].as_int == INVALID_MV)
1474 if (has_second_rf && this_mode == NEWMV &&
1475 mbmi->interp_filter == EIGHTTAP) {
1476 // adjust src pointers
1478 if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
1480 joint_motion_search(cpi, x, bsize, frame_mv[this_mode],
1481 mi_row, mi_col, seg_mvs[i],
1483 seg_mvs[i][mbmi->ref_frame[0]].as_int =
1484 frame_mv[this_mode][mbmi->ref_frame[0]].as_int;
1485 seg_mvs[i][mbmi->ref_frame[1]].as_int =
1486 frame_mv[this_mode][mbmi->ref_frame[1]].as_int;
1488 // restore src pointers
1489 mi_buf_restore(x, orig_src, orig_pre);
1492 bsi->rdstat[i][mode_idx].brate =
1493 set_and_cost_bmi_mvs(cpi, xd, i, this_mode, mode_mv[this_mode],
1494 frame_mv, seg_mvs[i], bsi->ref_mv,
1495 x->nmvjointcost, x->mvcost);
1497 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1498 bsi->rdstat[i][mode_idx].mvs[ref].as_int =
1499 mode_mv[this_mode][ref].as_int;
1500 if (num_4x4_blocks_wide > 1)
1501 bsi->rdstat[i + 1][mode_idx].mvs[ref].as_int =
1502 mode_mv[this_mode][ref].as_int;
1503 if (num_4x4_blocks_high > 1)
1504 bsi->rdstat[i + 2][mode_idx].mvs[ref].as_int =
1505 mode_mv[this_mode][ref].as_int;
1508 // Trap vectors that reach beyond the UMV borders
1509 if (mv_check_bounds(x, &mode_mv[this_mode][0].as_mv) ||
1511 mv_check_bounds(x, &mode_mv[this_mode][1].as_mv)))
1514 if (filter_idx > 0) {
1515 BEST_SEG_INFO *ref_bsi = bsi_buf;
1519 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1520 subpelmv |= mv_has_subpel(&mode_mv[this_mode][ref].as_mv);
1521 have_ref &= mode_mv[this_mode][ref].as_int ==
1522 ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
1525 if (filter_idx > 1 && !subpelmv && !have_ref) {
1526 ref_bsi = bsi_buf + 1;
1528 for (ref = 0; ref < 1 + has_second_rf; ++ref)
1529 have_ref &= mode_mv[this_mode][ref].as_int ==
1530 ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
1533 if (!subpelmv && have_ref &&
1534 ref_bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
1535 vpx_memcpy(&bsi->rdstat[i][mode_idx], &ref_bsi->rdstat[i][mode_idx],
1536 sizeof(SEG_RDSTAT));
1537 if (num_4x4_blocks_wide > 1)
1538 bsi->rdstat[i + 1][mode_idx].eobs =
1539 ref_bsi->rdstat[i + 1][mode_idx].eobs;
1540 if (num_4x4_blocks_high > 1)
1541 bsi->rdstat[i + 2][mode_idx].eobs =
1542 ref_bsi->rdstat[i + 2][mode_idx].eobs;
1544 if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
1545 mode_selected = this_mode;
1546 best_rd = bsi->rdstat[i][mode_idx].brdcost;
1552 bsi->rdstat[i][mode_idx].brdcost =
1553 encode_inter_mb_segment(cpi, x,
1554 bsi->segment_rd - this_segment_rd, i,
1555 &bsi->rdstat[i][mode_idx].byrate,
1556 &bsi->rdstat[i][mode_idx].bdist,
1557 &bsi->rdstat[i][mode_idx].bsse,
1558 bsi->rdstat[i][mode_idx].ta,
1559 bsi->rdstat[i][mode_idx].tl,
1561 if (bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
1562 bsi->rdstat[i][mode_idx].brdcost += RDCOST(x->rdmult, x->rddiv,
1563 bsi->rdstat[i][mode_idx].brate, 0);
1564 bsi->rdstat[i][mode_idx].brate += bsi->rdstat[i][mode_idx].byrate;
1565 bsi->rdstat[i][mode_idx].eobs = p->eobs[i];
1566 if (num_4x4_blocks_wide > 1)
1567 bsi->rdstat[i + 1][mode_idx].eobs = p->eobs[i + 1];
1568 if (num_4x4_blocks_high > 1)
1569 bsi->rdstat[i + 2][mode_idx].eobs = p->eobs[i + 2];
1572 if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
1573 mode_selected = this_mode;
1574 best_rd = bsi->rdstat[i][mode_idx].brdcost;
1576 } /*for each 4x4 mode*/
1578 if (best_rd == INT64_MAX) {
1580 for (iy = i + 1; iy < 4; ++iy)
1581 for (midx = 0; midx < INTER_MODES; ++midx)
1582 bsi->rdstat[iy][midx].brdcost = INT64_MAX;
1583 bsi->segment_rd = INT64_MAX;
1587 mode_idx = INTER_OFFSET(mode_selected);
1588 vpx_memcpy(t_above, bsi->rdstat[i][mode_idx].ta, sizeof(t_above));
1589 vpx_memcpy(t_left, bsi->rdstat[i][mode_idx].tl, sizeof(t_left));
1591 set_and_cost_bmi_mvs(cpi, xd, i, mode_selected, mode_mv[mode_selected],
1592 frame_mv, seg_mvs[i], bsi->ref_mv, x->nmvjointcost,
1595 br += bsi->rdstat[i][mode_idx].brate;
1596 bd += bsi->rdstat[i][mode_idx].bdist;
1597 block_sse += bsi->rdstat[i][mode_idx].bsse;
1598 segmentyrate += bsi->rdstat[i][mode_idx].byrate;
1599 this_segment_rd += bsi->rdstat[i][mode_idx].brdcost;
1601 if (this_segment_rd > bsi->segment_rd) {
1603 for (iy = i + 1; iy < 4; ++iy)
1604 for (midx = 0; midx < INTER_MODES; ++midx)
1605 bsi->rdstat[iy][midx].brdcost = INT64_MAX;
1606 bsi->segment_rd = INT64_MAX;
1610 } /* for each label */
1614 bsi->segment_yrate = segmentyrate;
1615 bsi->segment_rd = this_segment_rd;
1616 bsi->sse = block_sse;
1618 // update the coding decisions
1619 for (k = 0; k < 4; ++k)
1620 bsi->modes[k] = mi->bmi[k].as_mode;
1622 if (bsi->segment_rd > best_rd)
1624 /* set it to the best */
1625 for (i = 0; i < 4; i++) {
1626 mode_idx = INTER_OFFSET(bsi->modes[i]);
1627 mi->bmi[i].as_mv[0].as_int = bsi->rdstat[i][mode_idx].mvs[0].as_int;
1628 if (has_second_ref(mbmi))
1629 mi->bmi[i].as_mv[1].as_int = bsi->rdstat[i][mode_idx].mvs[1].as_int;
1630 x->plane[0].eobs[i] = bsi->rdstat[i][mode_idx].eobs;
1631 mi->bmi[i].as_mode = bsi->modes[i];
1635 * used to set mbmi->mv.as_int
1637 *returntotrate = bsi->r;
1638 *returndistortion = bsi->d;
1639 *returnyrate = bsi->segment_yrate;
1640 *skippable = vp9_is_skippable_in_plane(x, BLOCK_8X8, 0);
1642 mbmi->mode = bsi->modes[3];
1644 return bsi->segment_rd;
1647 static void estimate_ref_frame_costs(const VP9_COMMON *cm,
1648 const MACROBLOCKD *xd,
1650 unsigned int *ref_costs_single,
1651 unsigned int *ref_costs_comp,
1652 vp9_prob *comp_mode_p) {
1653 int seg_ref_active = vp9_segfeature_active(&cm->seg, segment_id,
1655 if (seg_ref_active) {
1656 vpx_memset(ref_costs_single, 0, MAX_REF_FRAMES * sizeof(*ref_costs_single));
1657 vpx_memset(ref_costs_comp, 0, MAX_REF_FRAMES * sizeof(*ref_costs_comp));
1660 vp9_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
1661 vp9_prob comp_inter_p = 128;
1663 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
1664 comp_inter_p = vp9_get_reference_mode_prob(cm, xd);
1665 *comp_mode_p = comp_inter_p;
1670 ref_costs_single[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
1672 if (cm->reference_mode != COMPOUND_REFERENCE) {
1673 vp9_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
1674 vp9_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
1675 unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
1677 if (cm->reference_mode == REFERENCE_MODE_SELECT)
1678 base_cost += vp9_cost_bit(comp_inter_p, 0);
1680 ref_costs_single[LAST_FRAME] = ref_costs_single[GOLDEN_FRAME] =
1681 ref_costs_single[ALTREF_FRAME] = base_cost;
1682 ref_costs_single[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
1683 ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1684 ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1685 ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
1686 ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
1688 ref_costs_single[LAST_FRAME] = 512;
1689 ref_costs_single[GOLDEN_FRAME] = 512;
1690 ref_costs_single[ALTREF_FRAME] = 512;
1692 if (cm->reference_mode != SINGLE_REFERENCE) {
1693 vp9_prob ref_comp_p = vp9_get_pred_prob_comp_ref_p(cm, xd);
1694 unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
1696 if (cm->reference_mode == REFERENCE_MODE_SELECT)
1697 base_cost += vp9_cost_bit(comp_inter_p, 1);
1699 ref_costs_comp[LAST_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 0);
1700 ref_costs_comp[GOLDEN_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 1);
1702 ref_costs_comp[LAST_FRAME] = 512;
1703 ref_costs_comp[GOLDEN_FRAME] = 512;
1708 static void store_coding_context(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
1710 int64_t comp_pred_diff[REFERENCE_MODES],
1711 const int64_t tx_size_diff[TX_MODES],
1712 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS],
1714 MACROBLOCKD *const xd = &x->e_mbd;
1716 // Take a snapshot of the coding context so it can be
1717 // restored if we decide to encode this way
1718 ctx->skip = x->skip;
1719 ctx->skippable = skippable;
1720 ctx->best_mode_index = mode_index;
1721 ctx->mic = *xd->mi[0];
1722 ctx->single_pred_diff = (int)comp_pred_diff[SINGLE_REFERENCE];
1723 ctx->comp_pred_diff = (int)comp_pred_diff[COMPOUND_REFERENCE];
1724 ctx->hybrid_pred_diff = (int)comp_pred_diff[REFERENCE_MODE_SELECT];
1726 vpx_memcpy(ctx->tx_rd_diff, tx_size_diff, sizeof(ctx->tx_rd_diff));
1727 vpx_memcpy(ctx->best_filter_diff, best_filter_diff,
1728 sizeof(*best_filter_diff) * SWITCHABLE_FILTER_CONTEXTS);
1731 static void setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
1732 const TileInfo *const tile,
1733 MV_REFERENCE_FRAME ref_frame,
1734 BLOCK_SIZE block_size,
1735 int mi_row, int mi_col,
1736 int_mv frame_nearest_mv[MAX_REF_FRAMES],
1737 int_mv frame_near_mv[MAX_REF_FRAMES],
1738 struct buf_2d yv12_mb[4][MAX_MB_PLANE]) {
1739 const VP9_COMMON *cm = &cpi->common;
1740 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
1741 MACROBLOCKD *const xd = &x->e_mbd;
1742 MODE_INFO *const mi = xd->mi[0];
1743 int_mv *const candidates = mi->mbmi.ref_mvs[ref_frame];
1744 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
1746 // TODO(jkoleszar): Is the UV buffer ever used here? If so, need to make this
1747 // use the UV scaling factors.
1748 vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
1750 // Gets an initial list of candidate vectors from neighbours and orders them
1751 vp9_find_mv_refs(cm, xd, tile, mi, ref_frame, candidates, mi_row, mi_col);
1753 // Candidate refinement carried out at encoder and decoder
1754 vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
1755 &frame_nearest_mv[ref_frame],
1756 &frame_near_mv[ref_frame]);
1758 // Further refinement that is encode side only to test the top few candidates
1759 // in full and choose the best as the centre point for subsequent searches.
1760 // The current implementation doesn't support scaling.
1761 if (!vp9_is_scaled(sf) && block_size >= BLOCK_8X8)
1762 vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride,
1763 ref_frame, block_size);
1766 static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
1768 int mi_row, int mi_col,
1769 int_mv *tmp_mv, int *rate_mv) {
1770 MACROBLOCKD *xd = &x->e_mbd;
1771 const VP9_COMMON *cm = &cpi->common;
1772 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
1773 struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
1774 int bestsme = INT_MAX;
1776 int sadpb = x->sadperbit16;
1778 int ref = mbmi->ref_frame[0];
1779 MV ref_mv = mbmi->ref_mvs[ref][0].as_mv;
1781 int tmp_col_min = x->mv_col_min;
1782 int tmp_col_max = x->mv_col_max;
1783 int tmp_row_min = x->mv_row_min;
1784 int tmp_row_max = x->mv_row_max;
1787 const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi,
1791 pred_mv[0] = mbmi->ref_mvs[ref][0].as_mv;
1792 pred_mv[1] = mbmi->ref_mvs[ref][1].as_mv;
1793 pred_mv[2] = x->pred_mv[ref];
1795 if (scaled_ref_frame) {
1797 // Swap out the reference frame for a version that's been scaled to
1798 // match the resolution of the current frame, allowing the existing
1799 // motion search code to be used without additional modifications.
1800 for (i = 0; i < MAX_MB_PLANE; i++)
1801 backup_yv12[i] = xd->plane[i].pre[0];
1803 vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
1806 vp9_set_mv_search_range(x, &ref_mv);
1808 // Work out the size of the first step in the mv step search.
1809 // 0 here is maximum length first step. 1 is MAX >> 1 etc.
1810 if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
1811 // Take wtd average of the step_params based on the last frame's
1812 // max mv magnitude and that based on the best ref mvs of the current
1813 // block for the given reference.
1814 step_param = (vp9_init_search_range(x->max_mv_context[ref]) +
1815 cpi->mv_step_param) / 2;
1817 step_param = cpi->mv_step_param;
1820 if (cpi->sf.adaptive_motion_search && bsize < BLOCK_64X64) {
1821 int boffset = 2 * (b_width_log2(BLOCK_64X64) - MIN(b_height_log2(bsize),
1822 b_width_log2(bsize)));
1823 step_param = MAX(step_param, boffset);
1826 if (cpi->sf.adaptive_motion_search) {
1827 int bwl = b_width_log2(bsize);
1828 int bhl = b_height_log2(bsize);
1830 int tlevel = x->pred_mv_sad[ref] >> (bwl + bhl + 4);
1835 for (i = LAST_FRAME; i <= ALTREF_FRAME && cm->show_frame; ++i) {
1836 if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
1837 x->pred_mv[ref].row = 0;
1838 x->pred_mv[ref].col = 0;
1839 tmp_mv->as_int = INVALID_MV;
1841 if (scaled_ref_frame) {
1843 for (i = 0; i < MAX_MB_PLANE; i++)
1844 xd->plane[i].pre[0] = backup_yv12[i];
1851 mvp_full = pred_mv[x->mv_best_ref_index[ref]];
1856 bestsme = vp9_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb,
1857 cond_sad_list(cpi, sad_list),
1858 &ref_mv, &tmp_mv->as_mv, INT_MAX, 1);
1860 x->mv_col_min = tmp_col_min;
1861 x->mv_col_max = tmp_col_max;
1862 x->mv_row_min = tmp_row_min;
1863 x->mv_row_max = tmp_row_max;
1865 if (bestsme < INT_MAX) {
1866 int dis; /* TODO: use dis in distortion calculation later. */
1867 cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
1868 cm->allow_high_precision_mv,
1870 &cpi->fn_ptr[bsize],
1871 cpi->sf.mv.subpel_force_stop,
1872 cpi->sf.mv.subpel_iters_per_step,
1873 cond_sad_list(cpi, sad_list),
1874 x->nmvjointcost, x->mvcost,
1875 &dis, &x->pred_sse[ref], NULL, 0, 0);
1877 *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv,
1878 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1880 if (cpi->sf.adaptive_motion_search)
1881 x->pred_mv[ref] = tmp_mv->as_mv;
1883 if (scaled_ref_frame) {
1885 for (i = 0; i < MAX_MB_PLANE; i++)
1886 xd->plane[i].pre[0] = backup_yv12[i];
1890 static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
1893 int mi_row, int mi_col,
1894 int_mv single_newmv[MAX_REF_FRAMES],
1896 const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
1897 const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
1898 MACROBLOCKD *xd = &x->e_mbd;
1899 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
1900 const int refs[2] = { mbmi->ref_frame[0],
1901 mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1] };
1904 // Prediction buffer from second frame.
1905 uint8_t *second_pred = vpx_memalign(16, pw * ph * sizeof(uint8_t));
1906 const InterpKernel *kernel = vp9_get_interp_kernel(mbmi->interp_filter);
1908 // Do joint motion search in compound mode to get more accurate mv.
1909 struct buf_2d backup_yv12[2][MAX_MB_PLANE];
1910 struct buf_2d scaled_first_yv12 = xd->plane[0].pre[0];
1911 int last_besterr[2] = {INT_MAX, INT_MAX};
1912 const YV12_BUFFER_CONFIG *const scaled_ref_frame[2] = {
1913 vp9_get_scaled_ref_frame(cpi, mbmi->ref_frame[0]),
1914 vp9_get_scaled_ref_frame(cpi, mbmi->ref_frame[1])
1917 for (ref = 0; ref < 2; ++ref) {
1918 ref_mv[ref] = mbmi->ref_mvs[refs[ref]][0];
1920 if (scaled_ref_frame[ref]) {
1922 // Swap out the reference frame for a version that's been scaled to
1923 // match the resolution of the current frame, allowing the existing
1924 // motion search code to be used without additional modifications.
1925 for (i = 0; i < MAX_MB_PLANE; i++)
1926 backup_yv12[ref][i] = xd->plane[i].pre[ref];
1927 vp9_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
1931 frame_mv[refs[ref]].as_int = single_newmv[refs[ref]].as_int;
1934 // Allow joint search multiple times iteratively for each ref frame
1935 // and break out the search loop if it couldn't find better mv.
1936 for (ite = 0; ite < 4; ite++) {
1937 struct buf_2d ref_yv12[2];
1938 int bestsme = INT_MAX;
1939 int sadpb = x->sadperbit16;
1941 int search_range = 3;
1943 int tmp_col_min = x->mv_col_min;
1944 int tmp_col_max = x->mv_col_max;
1945 int tmp_row_min = x->mv_row_min;
1946 int tmp_row_max = x->mv_row_max;
1949 // Initialized here because of compiler problem in Visual Studio.
1950 ref_yv12[0] = xd->plane[0].pre[0];
1951 ref_yv12[1] = xd->plane[0].pre[1];
1953 // Get pred block from second frame.
1954 vp9_build_inter_predictor(ref_yv12[!id].buf,
1955 ref_yv12[!id].stride,
1957 &frame_mv[refs[!id]].as_mv,
1958 &xd->block_refs[!id]->sf,
1960 kernel, MV_PRECISION_Q3,
1961 mi_col * MI_SIZE, mi_row * MI_SIZE);
1963 // Compound motion search on first ref frame.
1965 xd->plane[0].pre[0] = ref_yv12[id];
1966 vp9_set_mv_search_range(x, &ref_mv[id].as_mv);
1968 // Use mv result from single mode as mvp.
1969 tmp_mv = frame_mv[refs[id]].as_mv;
1974 // Small-range full-pixel motion search
1975 bestsme = vp9_refining_search_8p_c(x, &tmp_mv, sadpb,
1977 &cpi->fn_ptr[bsize],
1978 &ref_mv[id].as_mv, second_pred);
1979 if (bestsme < INT_MAX)
1980 bestsme = vp9_get_mvpred_av_var(x, &tmp_mv, &ref_mv[id].as_mv,
1981 second_pred, &cpi->fn_ptr[bsize], 1);
1983 x->mv_col_min = tmp_col_min;
1984 x->mv_col_max = tmp_col_max;
1985 x->mv_row_min = tmp_row_min;
1986 x->mv_row_max = tmp_row_max;
1988 if (bestsme < INT_MAX) {
1989 int dis; /* TODO: use dis in distortion calculation later. */
1991 bestsme = cpi->find_fractional_mv_step(
1994 cpi->common.allow_high_precision_mv,
1996 &cpi->fn_ptr[bsize],
1997 0, cpi->sf.mv.subpel_iters_per_step,
1999 x->nmvjointcost, x->mvcost,
2000 &dis, &sse, second_pred,
2005 xd->plane[0].pre[0] = scaled_first_yv12;
2007 if (bestsme < last_besterr[id]) {
2008 frame_mv[refs[id]].as_mv = tmp_mv;
2009 last_besterr[id] = bestsme;
2017 for (ref = 0; ref < 2; ++ref) {
2018 if (scaled_ref_frame[ref]) {
2019 // restore the predictor
2021 for (i = 0; i < MAX_MB_PLANE; i++)
2022 xd->plane[i].pre[ref] = backup_yv12[ref][i];
2025 *rate_mv += vp9_mv_bit_cost(&frame_mv[refs[ref]].as_mv,
2026 &mbmi->ref_mvs[refs[ref]][0].as_mv,
2027 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2030 vpx_free(second_pred);
2033 static INLINE void restore_dst_buf(MACROBLOCKD *xd,
2034 uint8_t *orig_dst[MAX_MB_PLANE],
2035 int orig_dst_stride[MAX_MB_PLANE]) {
2037 for (i = 0; i < MAX_MB_PLANE; i++) {
2038 xd->plane[i].dst.buf = orig_dst[i];
2039 xd->plane[i].dst.stride = orig_dst_stride[i];
2043 static void rd_encode_breakout_test(VP9_COMP *cpi, MACROBLOCK *x,
2044 BLOCK_SIZE bsize, int *rate2,
2045 int64_t *distortion, int64_t *distortion_uv,
2046 int *disable_skip) {
2047 VP9_COMMON *cm = &cpi->common;
2048 MACROBLOCKD *xd = &x->e_mbd;
2049 const BLOCK_SIZE y_size = get_plane_block_size(bsize, &xd->plane[0]);
2050 const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
2051 unsigned int var, sse;
2052 // Skipping threshold for ac.
2053 unsigned int thresh_ac;
2054 // Skipping threshold for dc
2055 unsigned int thresh_dc;
2057 var = cpi->fn_ptr[y_size].vf(x->plane[0].src.buf, x->plane[0].src.stride,
2058 xd->plane[0].dst.buf,
2059 xd->plane[0].dst.stride, &sse);
2061 if (x->encode_breakout > 0) {
2062 // Set a maximum for threshold to avoid big PSNR loss in low bitrate
2063 // case. Use extreme low threshold for static frames to limit skipping.
2064 const unsigned int max_thresh = (cpi->allow_encode_breakout ==
2065 ENCODE_BREAKOUT_LIMITED) ? 128 : 36000;
2066 // The encode_breakout input
2067 const unsigned int min_thresh =
2068 MIN(((unsigned int)x->encode_breakout << 4), max_thresh);
2070 // Calculate threshold according to dequant value.
2071 thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) / 9;
2072 thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);
2074 // Adjust threshold according to partition size.
2075 thresh_ac >>= 8 - (b_width_log2(bsize) +
2076 b_height_log2(bsize));
2077 thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
2083 // Y skipping condition checking
2084 if (sse < thresh_ac || sse == 0) {
2085 // dc skipping checking
2086 if ((sse - var) < thresh_dc || sse == var) {
2087 unsigned int sse_u, sse_v;
2088 unsigned int var_u, var_v;
2090 var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf,
2091 x->plane[1].src.stride,
2092 xd->plane[1].dst.buf,
2093 xd->plane[1].dst.stride, &sse_u);
2095 // U skipping condition checking
2096 if ((sse_u * 4 < thresh_ac || sse_u == 0) &&
2097 (sse_u - var_u < thresh_dc || sse_u == var_u)) {
2098 var_v = cpi->fn_ptr[uv_size].vf(x->plane[2].src.buf,
2099 x->plane[2].src.stride,
2100 xd->plane[2].dst.buf,
2101 xd->plane[2].dst.stride, &sse_v);
2103 // V skipping condition checking
2104 if ((sse_v * 4 < thresh_ac || sse_v == 0) &&
2105 (sse_v - var_v < thresh_dc || sse_v == var_v)) {
2108 // The cost of skip bit needs to be added.
2109 *rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2111 // Scaling factor for SSE from spatial domain to frequency domain
2112 // is 16. Adjust distortion accordingly.
2113 *distortion_uv = (sse_u + sse_v) << 4;
2114 *distortion = (sse << 4) + *distortion_uv;
2123 static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
2125 int64_t txfm_cache[],
2126 int *rate2, int64_t *distortion,
2128 int *rate_y, int64_t *distortion_y,
2129 int *rate_uv, int64_t *distortion_uv,
2131 int_mv (*mode_mv)[MAX_REF_FRAMES],
2132 int mi_row, int mi_col,
2133 int_mv single_newmv[MAX_REF_FRAMES],
2134 INTERP_FILTER (*single_filter)[MAX_REF_FRAMES],
2135 int (*single_skippable)[MAX_REF_FRAMES],
2137 const int64_t ref_best_rd) {
2138 VP9_COMMON *cm = &cpi->common;
2139 RD_OPT *rd_opt = &cpi->rd;
2140 MACROBLOCKD *xd = &x->e_mbd;
2141 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
2142 const int is_comp_pred = has_second_ref(mbmi);
2143 const int this_mode = mbmi->mode;
2144 int_mv *frame_mv = mode_mv[this_mode];
2146 int refs[2] = { mbmi->ref_frame[0],
2147 (mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1]) };
2149 int64_t this_rd = 0;
2150 DECLARE_ALIGNED_ARRAY(16, uint8_t, tmp_buf, MAX_MB_PLANE * 64 * 64);
2151 int pred_exists = 0;
2153 int64_t rd, tmp_rd, best_rd = INT64_MAX;
2154 int best_needs_copy = 0;
2155 uint8_t *orig_dst[MAX_MB_PLANE];
2156 int orig_dst_stride[MAX_MB_PLANE];
2158 INTERP_FILTER best_filter = SWITCHABLE;
2159 uint8_t skip_txfm[MAX_MB_PLANE << 2] = {0};
2160 int64_t bsse[MAX_MB_PLANE << 2] = {0};
2162 int bsl = mi_width_log2_lookup[bsize];
2163 int pred_filter_search = cpi->sf.cb_pred_filter_search ?
2164 (((mi_row + mi_col) >> bsl) +
2165 get_chessboard_index(cm->current_video_frame)) & 0x1 : 0;
2167 if (pred_filter_search) {
2168 INTERP_FILTER af = SWITCHABLE, lf = SWITCHABLE;
2169 if (xd->up_available)
2170 af = xd->mi[-xd->mi_stride]->mbmi.interp_filter;
2171 if (xd->left_available)
2172 lf = xd->mi[-1]->mbmi.interp_filter;
2174 if ((this_mode != NEWMV) || (af == lf))
2179 if (frame_mv[refs[0]].as_int == INVALID_MV ||
2180 frame_mv[refs[1]].as_int == INVALID_MV)
2183 if (cpi->sf.adaptive_mode_search) {
2184 if (single_filter[this_mode][refs[0]] ==
2185 single_filter[this_mode][refs[1]])
2186 best_filter = single_filter[this_mode][refs[0]];
2190 if (this_mode == NEWMV) {
2193 // Initialize mv using single prediction mode result.
2194 frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int;
2195 frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int;
2197 if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
2198 joint_motion_search(cpi, x, bsize, frame_mv,
2199 mi_row, mi_col, single_newmv, &rate_mv);
2201 rate_mv = vp9_mv_bit_cost(&frame_mv[refs[0]].as_mv,
2202 &mbmi->ref_mvs[refs[0]][0].as_mv,
2203 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2204 rate_mv += vp9_mv_bit_cost(&frame_mv[refs[1]].as_mv,
2205 &mbmi->ref_mvs[refs[1]][0].as_mv,
2206 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2211 single_motion_search(cpi, x, bsize, mi_row, mi_col,
2213 if (tmp_mv.as_int == INVALID_MV)
2216 frame_mv[refs[0]].as_int =
2217 xd->mi[0]->bmi[0].as_mv[0].as_int = tmp_mv.as_int;
2218 single_newmv[refs[0]].as_int = tmp_mv.as_int;
2222 for (i = 0; i < is_comp_pred + 1; ++i) {
2223 cur_mv[i] = frame_mv[refs[i]];
2224 // Clip "next_nearest" so that it does not extend to far out of image
2225 if (this_mode != NEWMV)
2226 clamp_mv2(&cur_mv[i].as_mv, xd);
2228 if (mv_check_bounds(x, &cur_mv[i].as_mv))
2230 mbmi->mv[i].as_int = cur_mv[i].as_int;
2233 // do first prediction into the destination buffer. Do the next
2234 // prediction into a temporary buffer. Then keep track of which one
2235 // of these currently holds the best predictor, and use the other
2236 // one for future predictions. In the end, copy from tmp_buf to
2237 // dst if necessary.
2238 for (i = 0; i < MAX_MB_PLANE; i++) {
2239 orig_dst[i] = xd->plane[i].dst.buf;
2240 orig_dst_stride[i] = xd->plane[i].dst.stride;
2243 /* We don't include the cost of the second reference here, because there
2244 * are only three options: Last/Golden, ARF/Last or Golden/ARF, or in other
2245 * words if you present them in that order, the second one is always known
2246 * if the first is known */
2247 *rate2 += cost_mv_ref(cpi, this_mode, mbmi->mode_context[refs[0]]);
2249 if (RDCOST(x->rdmult, x->rddiv, *rate2, 0) > ref_best_rd &&
2250 mbmi->mode != NEARESTMV)
2254 // Are all MVs integer pel for Y and UV
2255 intpel_mv = !mv_has_subpel(&mbmi->mv[0].as_mv);
2257 intpel_mv &= !mv_has_subpel(&mbmi->mv[1].as_mv);
2259 // Search for best switchable filter by checking the variance of
2260 // pred error irrespective of whether the filter will be used
2261 rd_opt->mask_filter = 0;
2262 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
2263 rd_opt->filter_cache[i] = INT64_MAX;
2265 if (cm->interp_filter != BILINEAR) {
2266 if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) {
2267 best_filter = EIGHTTAP;
2268 } else if (best_filter == SWITCHABLE) {
2270 int tmp_rate_sum = 0;
2271 int64_t tmp_dist_sum = 0;
2273 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
2276 mbmi->interp_filter = i;
2277 rs = vp9_get_switchable_rate(cpi);
2278 rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
2280 if (i > 0 && intpel_mv) {
2281 rd = RDCOST(x->rdmult, x->rddiv, tmp_rate_sum, tmp_dist_sum);
2282 rd_opt->filter_cache[i] = rd;
2283 rd_opt->filter_cache[SWITCHABLE_FILTERS] =
2284 MIN(rd_opt->filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2285 if (cm->interp_filter == SWITCHABLE)
2287 rd_opt->mask_filter = MAX(rd_opt->mask_filter, rd);
2290 int64_t dist_sum = 0;
2291 if (i > 0 && cpi->sf.adaptive_interp_filter_search &&
2292 (cpi->sf.interp_filter_search_mask & (1 << i))) {
2294 dist_sum = INT64_MAX;
2298 if ((cm->interp_filter == SWITCHABLE &&
2299 (!i || best_needs_copy)) ||
2300 (cm->interp_filter != SWITCHABLE &&
2301 (cm->interp_filter == mbmi->interp_filter ||
2302 (i == 0 && intpel_mv)))) {
2303 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2305 for (j = 0; j < MAX_MB_PLANE; j++) {
2306 xd->plane[j].dst.buf = tmp_buf + j * 64 * 64;
2307 xd->plane[j].dst.stride = 64;
2310 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2311 model_rd_for_sb(cpi, bsize, x, xd, &rate_sum, &dist_sum);
2313 rd = RDCOST(x->rdmult, x->rddiv, rate_sum, dist_sum);
2314 rd_opt->filter_cache[i] = rd;
2315 rd_opt->filter_cache[SWITCHABLE_FILTERS] =
2316 MIN(rd_opt->filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2317 if (cm->interp_filter == SWITCHABLE)
2319 rd_opt->mask_filter = MAX(rd_opt->mask_filter, rd);
2321 if (i == 0 && intpel_mv) {
2322 tmp_rate_sum = rate_sum;
2323 tmp_dist_sum = dist_sum;
2327 if (i == 0 && cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2328 if (rd / 2 > ref_best_rd) {
2329 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2333 newbest = i == 0 || rd < best_rd;
2337 best_filter = mbmi->interp_filter;
2338 if (cm->interp_filter == SWITCHABLE && i && !intpel_mv)
2339 best_needs_copy = !best_needs_copy;
2340 vpx_memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2341 vpx_memcpy(bsse, x->bsse, sizeof(bsse));
2344 if ((cm->interp_filter == SWITCHABLE && newbest) ||
2345 (cm->interp_filter != SWITCHABLE &&
2346 cm->interp_filter == mbmi->interp_filter)) {
2351 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2354 // Set the appropriate filter
2355 mbmi->interp_filter = cm->interp_filter != SWITCHABLE ?
2356 cm->interp_filter : best_filter;
2357 rs = cm->interp_filter == SWITCHABLE ? vp9_get_switchable_rate(cpi) : 0;
2360 if (best_needs_copy) {
2361 // again temporarily set the buffers to local memory to prevent a memcpy
2362 for (i = 0; i < MAX_MB_PLANE; i++) {
2363 xd->plane[i].dst.buf = tmp_buf + i * 64 * 64;
2364 xd->plane[i].dst.stride = 64;
2367 rd = tmp_rd + RDCOST(x->rdmult, x->rddiv, rs, 0);
2371 // Handles the special case when a filter that is not in the
2372 // switchable list (ex. bilinear) is indicated at the frame level, or
2373 // skip condition holds.
2374 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2375 model_rd_for_sb(cpi, bsize, x, xd, &tmp_rate, &tmp_dist);
2376 rd = RDCOST(x->rdmult, x->rddiv, rs + tmp_rate, tmp_dist);
2377 vpx_memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2378 vpx_memcpy(bsse, x->bsse, sizeof(bsse));
2382 single_filter[this_mode][refs[0]] = mbmi->interp_filter;
2384 if (cpi->sf.adaptive_mode_search)
2386 if (single_skippable[this_mode][refs[0]] &&
2387 single_skippable[this_mode][refs[1]])
2388 vpx_memset(skip_txfm, 1, sizeof(skip_txfm));
2390 if (cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2391 // if current pred_error modeled rd is substantially more than the best
2392 // so far, do not bother doing full rd
2393 if (rd / 2 > ref_best_rd) {
2394 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2399 if (cm->interp_filter == SWITCHABLE)
2402 if (!is_comp_pred) {
2403 if (cpi->allow_encode_breakout)
2404 rd_encode_breakout_test(cpi, x, bsize, rate2, distortion, distortion_uv,
2408 vpx_memcpy(x->skip_txfm, skip_txfm, sizeof(skip_txfm));
2409 vpx_memcpy(x->bsse, bsse, sizeof(bsse));
2412 int skippable_y, skippable_uv;
2413 int64_t sseuv = INT64_MAX;
2414 int64_t rdcosty = INT64_MAX;
2416 // Y cost and distortion
2417 vp9_subtract_plane(x, bsize, 0);
2418 super_block_yrd(cpi, x, rate_y, distortion_y, &skippable_y, psse,
2419 bsize, txfm_cache, ref_best_rd);
2421 if (*rate_y == INT_MAX) {
2423 *distortion = INT64_MAX;
2424 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2429 *distortion += *distortion_y;
2431 rdcosty = RDCOST(x->rdmult, x->rddiv, *rate2, *distortion);
2432 rdcosty = MIN(rdcosty, RDCOST(x->rdmult, x->rddiv, 0, *psse));
2434 super_block_uvrd(cpi, x, rate_uv, distortion_uv, &skippable_uv, &sseuv,
2435 bsize, ref_best_rd - rdcosty);
2436 if (*rate_uv == INT_MAX) {
2438 *distortion = INT64_MAX;
2439 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2445 *distortion += *distortion_uv;
2446 *skippable = skippable_y && skippable_uv;
2450 single_skippable[this_mode][refs[0]] = *skippable;
2452 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2453 return this_rd; // if 0, this will be re-calculated by caller
2456 void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
2457 int *returnrate, int64_t *returndist,
2459 PICK_MODE_CONTEXT *ctx, int64_t best_rd) {
2460 VP9_COMMON *const cm = &cpi->common;
2461 MACROBLOCKD *const xd = &x->e_mbd;
2462 struct macroblockd_plane *const pd = xd->plane;
2463 int rate_y = 0, rate_uv = 0, rate_y_tokenonly = 0, rate_uv_tokenonly = 0;
2464 int y_skip = 0, uv_skip = 0;
2465 int64_t dist_y = 0, dist_uv = 0, tx_cache[TX_MODES] = { 0 };
2466 TX_SIZE max_uv_tx_size;
2469 xd->mi[0]->mbmi.ref_frame[0] = INTRA_FRAME;
2471 if (bsize >= BLOCK_8X8) {
2472 if (rd_pick_intra_sby_mode(cpi, x, &rate_y, &rate_y_tokenonly,
2473 &dist_y, &y_skip, bsize, tx_cache,
2474 best_rd) >= best_rd) {
2475 *returnrate = INT_MAX;
2478 max_uv_tx_size = get_uv_tx_size_impl(xd->mi[0]->mbmi.tx_size, bsize,
2479 pd[1].subsampling_x,
2480 pd[1].subsampling_y);
2481 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly,
2482 &dist_uv, &uv_skip, bsize, max_uv_tx_size);
2485 if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate_y, &rate_y_tokenonly,
2486 &dist_y, best_rd) >= best_rd) {
2487 *returnrate = INT_MAX;
2490 max_uv_tx_size = get_uv_tx_size_impl(xd->mi[0]->mbmi.tx_size, bsize,
2491 pd[1].subsampling_x,
2492 pd[1].subsampling_y);
2493 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly,
2494 &dist_uv, &uv_skip, BLOCK_8X8, max_uv_tx_size);
2497 if (y_skip && uv_skip) {
2498 *returnrate = rate_y + rate_uv - rate_y_tokenonly - rate_uv_tokenonly +
2499 vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2500 *returndist = dist_y + dist_uv;
2501 vp9_zero(ctx->tx_rd_diff);
2504 *returnrate = rate_y + rate_uv + vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
2505 *returndist = dist_y + dist_uv;
2506 if (cpi->sf.tx_size_search_method == USE_FULL_RD)
2507 for (i = 0; i < TX_MODES; i++) {
2508 if (tx_cache[i] < INT64_MAX && tx_cache[cm->tx_mode] < INT64_MAX)
2509 ctx->tx_rd_diff[i] = tx_cache[i] - tx_cache[cm->tx_mode];
2511 ctx->tx_rd_diff[i] = 0;
2515 ctx->mic = *xd->mi[0];
2518 // Updating rd_thresh_freq_fact[] here means that the different
2519 // partition/block sizes are handled independently based on the best
2520 // choice for the current partition. It may well be better to keep a scaled
2521 // best rd so far value and update rd_thresh_freq_fact based on the mode/size
2522 // combination that wins out.
2523 static void update_rd_thresh_fact(VP9_COMP *cpi, int bsize,
2524 int best_mode_index) {
2525 if (cpi->sf.adaptive_rd_thresh > 0) {
2526 const int top_mode = bsize < BLOCK_8X8 ? MAX_REFS : MAX_MODES;
2528 for (mode = 0; mode < top_mode; ++mode) {
2529 int *const fact = &cpi->rd.thresh_freq_fact[bsize][mode];
2531 if (mode == best_mode_index) {
2532 *fact -= (*fact >> 3);
2534 *fact = MIN(*fact + RD_THRESH_INC,
2535 cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
2541 int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
2542 const TileInfo *const tile,
2543 int mi_row, int mi_col,
2545 int64_t *returndistortion,
2547 PICK_MODE_CONTEXT *ctx,
2548 int64_t best_rd_so_far) {
2549 VP9_COMMON *const cm = &cpi->common;
2550 RD_OPT *const rd_opt = &cpi->rd;
2551 MACROBLOCKD *const xd = &x->e_mbd;
2552 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
2553 const struct segmentation *const seg = &cm->seg;
2554 struct macroblockd_plane *const pd = xd->plane;
2555 PREDICTION_MODE this_mode;
2556 MV_REFERENCE_FRAME ref_frame, second_ref_frame;
2557 unsigned char segment_id = mbmi->segment_id;
2558 int comp_pred, i, k;
2559 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
2560 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
2561 int_mv single_newmv[MAX_REF_FRAMES] = { { 0 } };
2562 INTERP_FILTER single_inter_filter[MB_MODE_COUNT][MAX_REF_FRAMES];
2563 int single_skippable[MB_MODE_COUNT][MAX_REF_FRAMES];
2564 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
2566 int64_t best_rd = best_rd_so_far;
2567 int64_t best_tx_rd[TX_MODES];
2568 int64_t best_tx_diff[TX_MODES];
2569 int64_t best_pred_diff[REFERENCE_MODES];
2570 int64_t best_pred_rd[REFERENCE_MODES];
2571 int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
2572 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
2573 MB_MODE_INFO best_mbmode;
2574 int best_mode_skippable = 0;
2575 int mode_index, best_mode_index = -1;
2576 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
2577 vp9_prob comp_mode_p;
2578 int64_t best_intra_rd = INT64_MAX;
2579 unsigned int best_pred_sse = UINT_MAX;
2580 PREDICTION_MODE best_intra_mode = DC_PRED;
2581 int rate_uv_intra[TX_SIZES], rate_uv_tokenonly[TX_SIZES];
2582 int64_t dist_uv[TX_SIZES];
2583 int skip_uv[TX_SIZES];
2584 PREDICTION_MODE mode_uv[TX_SIZES];
2585 int intra_cost_penalty = 20 * vp9_dc_quant(cm->base_qindex, cm->y_dc_delta_q);
2587 uint8_t ref_frame_skip_mask[2] = { 0 };
2588 uint16_t mode_skip_mask[MAX_REF_FRAMES] = { 0 };
2589 int mode_skip_start = cpi->sf.mode_skip_start + 1;
2590 const int *const rd_threshes = rd_opt->threshes[segment_id][bsize];
2591 const int *const rd_thresh_freq_fact = rd_opt->thresh_freq_fact[bsize];
2592 const int mode_search_skip_flags = cpi->sf.mode_search_skip_flags;
2593 vp9_zero(best_mbmode);
2594 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
2596 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
2599 for (i = 0; i < REFERENCE_MODES; ++i)
2600 best_pred_rd[i] = INT64_MAX;
2601 for (i = 0; i < TX_MODES; i++)
2602 best_tx_rd[i] = INT64_MAX;
2603 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
2604 best_filter_rd[i] = INT64_MAX;
2605 for (i = 0; i < TX_SIZES; i++)
2606 rate_uv_intra[i] = INT_MAX;
2607 for (i = 0; i < MAX_REF_FRAMES; ++i)
2608 x->pred_sse[i] = INT_MAX;
2609 for (i = 0; i < MB_MODE_COUNT; ++i) {
2610 for (k = 0; k < MAX_REF_FRAMES; ++k) {
2611 single_inter_filter[i][k] = SWITCHABLE;
2612 single_skippable[i][k] = 0;
2616 *returnrate = INT_MAX;
2618 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
2619 x->pred_mv_sad[ref_frame] = INT_MAX;
2620 if (cpi->ref_frame_flags & flag_list[ref_frame]) {
2621 setup_buffer_inter(cpi, x, tile, ref_frame, bsize, mi_row, mi_col,
2622 frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
2624 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
2625 frame_mv[ZEROMV][ref_frame].as_int = 0;
2628 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
2629 if (!(cpi->ref_frame_flags & flag_list[ref_frame])) {
2630 // Skip checking missing references in both single and compound reference
2631 // modes. Note that a mode will be skipped iff both reference frames
2633 ref_frame_skip_mask[0] |= (1 << ref_frame);
2634 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
2635 } else if (cpi->sf.reference_masking) {
2636 for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
2637 // Skip fixed mv modes for poor references
2638 if ((x->pred_mv_sad[ref_frame] >> 2) > x->pred_mv_sad[i]) {
2639 mode_skip_mask[ref_frame] |= INTER_NEAREST_NEAR_ZERO;
2644 // If the segment reference frame feature is enabled....
2645 // then do nothing if the current ref frame is not allowed..
2646 if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
2647 vp9_get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
2648 ref_frame_skip_mask[0] |= (1 << ref_frame);
2649 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
2653 // Disable this drop out case if the ref frame
2654 // segment level feature is enabled for this segment. This is to
2655 // prevent the possibility that we end up unable to pick any mode.
2656 if (!vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
2657 // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
2658 // unless ARNR filtering is enabled in which case we want
2659 // an unfiltered alternative. We allow near/nearest as well
2660 // because they may result in zero-zero MVs but be cheaper.
2661 if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) {
2662 ref_frame_skip_mask[0] = (1 << LAST_FRAME) | (1 << GOLDEN_FRAME);
2663 ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
2664 mode_skip_mask[ALTREF_FRAME] = ~INTER_NEAREST_NEAR_ZERO;
2665 if (frame_mv[NEARMV][ALTREF_FRAME].as_int != 0)
2666 mode_skip_mask[ALTREF_FRAME] |= (1 << NEARMV);
2667 if (frame_mv[NEARESTMV][ALTREF_FRAME].as_int != 0)
2668 mode_skip_mask[ALTREF_FRAME] |= (1 << NEARESTMV);
2672 if (cpi->rc.is_src_frame_alt_ref) {
2673 if (cpi->sf.alt_ref_search_fp) {
2674 mode_skip_mask[ALTREF_FRAME] = 0;
2675 ref_frame_skip_mask[0] = ~(1 << ALTREF_FRAME);
2676 ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
2680 if (bsize > cpi->sf.max_intra_bsize) {
2681 ref_frame_skip_mask[0] |= (1 << INTRA_FRAME);
2682 ref_frame_skip_mask[1] |= (1 << INTRA_FRAME);
2685 mode_skip_mask[INTRA_FRAME] |=
2686 ~(cpi->sf.intra_y_mode_mask[max_txsize_lookup[bsize]]);
2688 for (mode_index = 0; mode_index < MAX_MODES; ++mode_index) {
2689 int mode_excluded = 0;
2690 int64_t this_rd = INT64_MAX;
2691 int disable_skip = 0;
2692 int compmode_cost = 0;
2693 int rate2 = 0, rate_y = 0, rate_uv = 0;
2694 int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
2696 int64_t tx_cache[TX_MODES];
2699 int64_t total_sse = INT64_MAX;
2702 this_mode = vp9_mode_order[mode_index].mode;
2703 ref_frame = vp9_mode_order[mode_index].ref_frame[0];
2704 second_ref_frame = vp9_mode_order[mode_index].ref_frame[1];
2706 // Look at the reference frame of the best mode so far and set the
2707 // skip mask to look at a subset of the remaining modes.
2708 if (mode_index == mode_skip_start && best_mode_index >= 0) {
2709 switch (best_mbmode.ref_frame[0]) {
2713 ref_frame_skip_mask[0] |= LAST_FRAME_MODE_MASK;
2714 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
2717 ref_frame_skip_mask[0] |= GOLDEN_FRAME_MODE_MASK;
2718 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
2721 ref_frame_skip_mask[0] |= ALT_REF_MODE_MASK;
2724 case MAX_REF_FRAMES:
2725 assert(0 && "Invalid Reference frame");
2730 if (ref_frame_skip_mask[0] & (1 << ref_frame) &&
2731 ref_frame_skip_mask[1] & (1 << MAX(0, second_ref_frame)))
2734 if (mode_skip_mask[ref_frame] & (1 << this_mode))
2737 // Test best rd so far against threshold for trying this mode.
2738 if (rd_less_than_thresh(best_rd, rd_threshes[mode_index],
2739 rd_thresh_freq_fact[mode_index]))
2742 if (cpi->sf.motion_field_mode_search) {
2743 const int mi_width = MIN(num_8x8_blocks_wide_lookup[bsize],
2744 tile->mi_col_end - mi_col);
2745 const int mi_height = MIN(num_8x8_blocks_high_lookup[bsize],
2746 tile->mi_row_end - mi_row);
2747 const int bsl = mi_width_log2(bsize);
2748 int cb_partition_search_ctrl = (((mi_row + mi_col) >> bsl)
2749 + get_chessboard_index(cm->current_video_frame)) & 0x1;
2750 MB_MODE_INFO *ref_mbmi;
2751 int const_motion = 1;
2752 int skip_ref_frame = !cb_partition_search_ctrl;
2753 MV_REFERENCE_FRAME rf = NONE;
2755 ref_mv.as_int = INVALID_MV;
2757 if ((mi_row - 1) >= tile->mi_row_start) {
2758 ref_mv = xd->mi[-xd->mi_stride]->mbmi.mv[0];
2759 rf = xd->mi[-xd->mi_stride]->mbmi.ref_frame[0];
2760 for (i = 0; i < mi_width; ++i) {
2761 ref_mbmi = &xd->mi[-xd->mi_stride + i]->mbmi;
2762 const_motion &= (ref_mv.as_int == ref_mbmi->mv[0].as_int) &&
2763 (ref_frame == ref_mbmi->ref_frame[0]);
2764 skip_ref_frame &= (rf == ref_mbmi->ref_frame[0]);
2768 if ((mi_col - 1) >= tile->mi_col_start) {
2769 if (ref_mv.as_int == INVALID_MV)
2770 ref_mv = xd->mi[-1]->mbmi.mv[0];
2772 rf = xd->mi[-1]->mbmi.ref_frame[0];
2773 for (i = 0; i < mi_height; ++i) {
2774 ref_mbmi = &xd->mi[i * xd->mi_stride - 1]->mbmi;
2775 const_motion &= (ref_mv.as_int == ref_mbmi->mv[0].as_int) &&
2776 (ref_frame == ref_mbmi->ref_frame[0]);
2777 skip_ref_frame &= (rf == ref_mbmi->ref_frame[0]);
2781 if (skip_ref_frame && this_mode != NEARESTMV && this_mode != NEWMV)
2782 if (rf > INTRA_FRAME)
2783 if (ref_frame != rf)
2787 if (this_mode == NEARMV || this_mode == ZEROMV)
2791 comp_pred = second_ref_frame > INTRA_FRAME;
2793 if (!cm->allow_comp_inter_inter)
2796 if ((mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
2797 best_mode_index >= 0 && best_mbmode.ref_frame[0] == INTRA_FRAME)
2800 mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
2802 if (ref_frame != INTRA_FRAME)
2803 mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
2806 if (ref_frame == INTRA_FRAME) {
2807 if (cpi->sf.adaptive_mode_search)
2808 if ((x->source_variance << num_pels_log2_lookup[bsize]) > best_pred_sse)
2811 if (this_mode != DC_PRED) {
2812 // Disable intra modes other than DC_PRED for blocks with low variance
2813 // Threshold for intra skipping based on source variance
2814 // TODO(debargha): Specialize the threshold for super block sizes
2815 const unsigned int skip_intra_var_thresh = 64;
2816 if ((mode_search_skip_flags & FLAG_SKIP_INTRA_LOWVAR) &&
2817 x->source_variance < skip_intra_var_thresh)
2819 // Only search the oblique modes if the best so far is
2820 // one of the neighboring directional modes
2821 if ((mode_search_skip_flags & FLAG_SKIP_INTRA_BESTINTER) &&
2822 (this_mode >= D45_PRED && this_mode <= TM_PRED)) {
2823 if (best_mode_index >= 0 &&
2824 best_mbmode.ref_frame[0] > INTRA_FRAME)
2827 if (mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
2828 if (conditional_skipintra(this_mode, best_intra_mode))
2833 const MV_REFERENCE_FRAME ref_frames[2] = {ref_frame, second_ref_frame};
2834 if (!check_best_zero_mv(cpi, mbmi->mode_context, frame_mv,
2835 this_mode, ref_frames))
2839 mbmi->mode = this_mode;
2840 mbmi->uv_mode = DC_PRED;
2841 mbmi->ref_frame[0] = ref_frame;
2842 mbmi->ref_frame[1] = second_ref_frame;
2843 // Evaluate all sub-pel filters irrespective of whether we can use
2844 // them for this frame.
2845 mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
2846 : cm->interp_filter;
2847 mbmi->mv[0].as_int = mbmi->mv[1].as_int = 0;
2850 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
2852 // Select prediction reference frames.
2853 for (i = 0; i < MAX_MB_PLANE; i++) {
2854 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
2856 xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
2859 for (i = 0; i < TX_MODES; ++i)
2860 tx_cache[i] = INT64_MAX;
2862 if (ref_frame == INTRA_FRAME) {
2864 super_block_yrd(cpi, x, &rate_y, &distortion_y, &skippable,
2865 NULL, bsize, tx_cache, best_rd);
2867 if (rate_y == INT_MAX)
2870 uv_tx = get_uv_tx_size_impl(mbmi->tx_size, bsize, pd[1].subsampling_x,
2871 pd[1].subsampling_y);
2872 if (rate_uv_intra[uv_tx] == INT_MAX) {
2873 choose_intra_uv_mode(cpi, ctx, bsize, uv_tx,
2874 &rate_uv_intra[uv_tx], &rate_uv_tokenonly[uv_tx],
2875 &dist_uv[uv_tx], &skip_uv[uv_tx], &mode_uv[uv_tx]);
2878 rate_uv = rate_uv_tokenonly[uv_tx];
2879 distortion_uv = dist_uv[uv_tx];
2880 skippable = skippable && skip_uv[uv_tx];
2881 mbmi->uv_mode = mode_uv[uv_tx];
2883 rate2 = rate_y + cpi->mbmode_cost[mbmi->mode] + rate_uv_intra[uv_tx];
2884 if (this_mode != DC_PRED && this_mode != TM_PRED)
2885 rate2 += intra_cost_penalty;
2886 distortion2 = distortion_y + distortion_uv;
2888 this_rd = handle_inter_mode(cpi, x, bsize,
2890 &rate2, &distortion2, &skippable,
2891 &rate_y, &distortion_y,
2892 &rate_uv, &distortion_uv,
2893 &disable_skip, frame_mv,
2895 single_newmv, single_inter_filter,
2896 single_skippable, &total_sse, best_rd);
2897 if (this_rd == INT64_MAX)
2900 compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
2902 if (cm->reference_mode == REFERENCE_MODE_SELECT)
2903 rate2 += compmode_cost;
2906 // Estimate the reference frame signaling cost and add it
2907 // to the rolling cost variable.
2909 rate2 += ref_costs_comp[ref_frame];
2911 rate2 += ref_costs_single[ref_frame];
2914 if (!disable_skip) {
2916 vp9_prob skip_prob = vp9_get_skip_prob(cm, xd);
2918 // Back out the coefficient coding costs
2919 rate2 -= (rate_y + rate_uv);
2920 // for best yrd calculation
2923 // Cost the skip mb case
2925 int prob_skip_cost = vp9_cost_bit(skip_prob, 1);
2926 rate2 += prob_skip_cost;
2928 } else if (ref_frame != INTRA_FRAME && !xd->lossless) {
2929 if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
2930 RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
2931 // Add in the cost of the no skip flag.
2932 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
2934 // FIXME(rbultje) make this work for splitmv also
2935 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2936 distortion2 = total_sse;
2937 assert(total_sse >= 0);
2938 rate2 -= (rate_y + rate_uv);
2944 // Add in the cost of the no skip flag.
2945 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
2948 // Calculate the final RD estimate for this mode.
2949 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
2952 if (ref_frame == INTRA_FRAME) {
2953 // Keep record of best intra rd
2954 if (this_rd < best_intra_rd) {
2955 best_intra_rd = this_rd;
2956 best_intra_mode = mbmi->mode;
2960 if (!disable_skip && ref_frame == INTRA_FRAME) {
2961 for (i = 0; i < REFERENCE_MODES; ++i)
2962 best_pred_rd[i] = MIN(best_pred_rd[i], this_rd);
2963 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
2964 best_filter_rd[i] = MIN(best_filter_rd[i], this_rd);
2967 // Did this mode help.. i.e. is it the new best mode
2968 if (this_rd < best_rd || x->skip) {
2969 int max_plane = MAX_MB_PLANE;
2970 if (!mode_excluded) {
2971 // Note index of best mode so far
2972 best_mode_index = mode_index;
2974 if (ref_frame == INTRA_FRAME) {
2975 /* required for left and above block mv */
2976 mbmi->mv[0].as_int = 0;
2979 best_pred_sse = x->pred_sse[ref_frame];
2982 *returnrate = rate2;
2983 *returndistortion = distortion2;
2985 best_mbmode = *mbmi;
2986 best_skip2 = this_skip2;
2987 best_mode_skippable = skippable;
2989 if (!x->select_tx_size)
2990 swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
2991 vpx_memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mbmi->tx_size],
2992 sizeof(uint8_t) * ctx->num_4x4_blk);
2994 // TODO(debargha): enhance this test with a better distortion prediction
2995 // based on qp, activity mask and history
2996 if ((mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
2997 (mode_index > MIN_EARLY_TERM_INDEX)) {
2998 const int qstep = xd->plane[0].dequant[1];
2999 // TODO(debargha): Enhance this by specializing for each mode_index
3001 if (x->source_variance < UINT_MAX) {
3002 const int var_adjust = (x->source_variance < 16);
3003 scale -= var_adjust;
3005 if (ref_frame > INTRA_FRAME &&
3006 distortion2 * scale < qstep * qstep) {
3013 /* keep record of best compound/single-only prediction */
3014 if (!disable_skip && ref_frame != INTRA_FRAME) {
3015 int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
3017 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
3018 single_rate = rate2 - compmode_cost;
3019 hybrid_rate = rate2;
3021 single_rate = rate2;
3022 hybrid_rate = rate2 + compmode_cost;
3025 single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
3026 hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
3029 if (single_rd < best_pred_rd[SINGLE_REFERENCE])
3030 best_pred_rd[SINGLE_REFERENCE] = single_rd;
3032 if (single_rd < best_pred_rd[COMPOUND_REFERENCE])
3033 best_pred_rd[COMPOUND_REFERENCE] = single_rd;
3035 if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
3036 best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
3038 /* keep record of best filter type */
3039 if (!mode_excluded && cm->interp_filter != BILINEAR) {
3040 int64_t ref = rd_opt->filter_cache[cm->interp_filter == SWITCHABLE ?
3041 SWITCHABLE_FILTERS : cm->interp_filter];
3043 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3045 if (ref == INT64_MAX)
3047 else if (rd_opt->filter_cache[i] == INT64_MAX)
3048 // when early termination is triggered, the encoder does not have
3049 // access to the rate-distortion cost. it only knows that the cost
3050 // should be above the maximum valid value. hence it takes the known
3051 // maximum plus an arbitrary constant as the rate-distortion cost.
3052 adj_rd = rd_opt->mask_filter - ref + 10;
3054 adj_rd = rd_opt->filter_cache[i] - ref;
3057 best_filter_rd[i] = MIN(best_filter_rd[i], adj_rd);
3062 /* keep record of best txfm size */
3063 if (bsize < BLOCK_32X32) {
3064 if (bsize < BLOCK_16X16)
3065 tx_cache[ALLOW_16X16] = tx_cache[ALLOW_8X8];
3067 tx_cache[ALLOW_32X32] = tx_cache[ALLOW_16X16];
3069 if (!mode_excluded && this_rd != INT64_MAX) {
3070 for (i = 0; i < TX_MODES && tx_cache[i] < INT64_MAX; i++) {
3071 int64_t adj_rd = INT64_MAX;
3072 adj_rd = this_rd + tx_cache[i] - tx_cache[cm->tx_mode];
3074 if (adj_rd < best_tx_rd[i])
3075 best_tx_rd[i] = adj_rd;
3082 if (x->skip && !comp_pred)
3086 // The inter modes' rate costs are not calculated precisely in some cases.
3087 // Therefore, sometimes, NEWMV is chosen instead of NEARESTMV, NEARMV, and
3088 // ZEROMV. Here, checks are added for those cases, and the mode decisions
3090 if (best_mbmode.mode == NEWMV) {
3091 const MV_REFERENCE_FRAME refs[2] = {best_mbmode.ref_frame[0],
3092 best_mbmode.ref_frame[1]};
3093 int comp_pred_mode = refs[1] > INTRA_FRAME;
3095 if (frame_mv[NEARESTMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3096 ((comp_pred_mode && frame_mv[NEARESTMV][refs[1]].as_int ==
3097 best_mbmode.mv[1].as_int) || !comp_pred_mode))
3098 best_mbmode.mode = NEARESTMV;
3099 else if (frame_mv[NEARMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3100 ((comp_pred_mode && frame_mv[NEARMV][refs[1]].as_int ==
3101 best_mbmode.mv[1].as_int) || !comp_pred_mode))
3102 best_mbmode.mode = NEARMV;
3103 else if (best_mbmode.mv[0].as_int == 0 &&
3104 ((comp_pred_mode && best_mbmode.mv[1].as_int == 0) || !comp_pred_mode))
3105 best_mbmode.mode = ZEROMV;
3108 if (best_mode_index < 0 || best_rd >= best_rd_so_far)
3111 // If we used an estimate for the uv intra rd in the loop above...
3112 if (cpi->sf.use_uv_intra_rd_estimate) {
3113 // Do Intra UV best rd mode selection if best mode choice above was intra.
3114 if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
3116 *mbmi = best_mbmode;
3117 uv_tx_size = get_uv_tx_size(mbmi, &xd->plane[1]);
3118 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra[uv_tx_size],
3119 &rate_uv_tokenonly[uv_tx_size],
3120 &dist_uv[uv_tx_size],
3121 &skip_uv[uv_tx_size],
3122 bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize,
3127 assert((cm->interp_filter == SWITCHABLE) ||
3128 (cm->interp_filter == best_mbmode.interp_filter) ||
3129 !is_inter_block(&best_mbmode));
3131 update_rd_thresh_fact(cpi, bsize, best_mode_index);
3134 *mbmi = best_mbmode;
3135 x->skip |= best_skip2;
3137 for (i = 0; i < REFERENCE_MODES; ++i) {
3138 if (best_pred_rd[i] == INT64_MAX)
3139 best_pred_diff[i] = INT_MIN;
3141 best_pred_diff[i] = best_rd - best_pred_rd[i];
3145 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3146 if (best_filter_rd[i] == INT64_MAX)
3147 best_filter_diff[i] = 0;
3149 best_filter_diff[i] = best_rd - best_filter_rd[i];
3151 if (cm->interp_filter == SWITCHABLE)
3152 assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
3153 for (i = 0; i < TX_MODES; i++) {
3154 if (best_tx_rd[i] == INT64_MAX)
3155 best_tx_diff[i] = 0;
3157 best_tx_diff[i] = best_rd - best_tx_rd[i];
3160 vp9_zero(best_filter_diff);
3161 vp9_zero(best_tx_diff);
3164 store_coding_context(x, ctx, best_mode_index, best_pred_diff,
3165 best_tx_diff, best_filter_diff, best_mode_skippable);
3170 int64_t vp9_rd_pick_inter_mode_sb_seg_skip(VP9_COMP *cpi, MACROBLOCK *x,
3172 int64_t *returndistortion,
3174 PICK_MODE_CONTEXT *ctx,
3175 int64_t best_rd_so_far) {
3176 VP9_COMMON *const cm = &cpi->common;
3177 RD_OPT *const rd_opt = &cpi->rd;
3178 MACROBLOCKD *const xd = &x->e_mbd;
3179 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
3180 unsigned char segment_id = mbmi->segment_id;
3181 const int comp_pred = 0;
3183 int64_t best_tx_diff[TX_MODES];
3184 int64_t best_pred_diff[REFERENCE_MODES];
3185 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3186 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3187 vp9_prob comp_mode_p;
3188 INTERP_FILTER best_filter = SWITCHABLE;
3189 int64_t this_rd = INT64_MAX;
3191 const int64_t distortion2 = 0;
3193 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3195 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3198 for (i = 0; i < MAX_REF_FRAMES; ++i)
3199 x->pred_sse[i] = INT_MAX;
3200 for (i = LAST_FRAME; i < MAX_REF_FRAMES; ++i)
3201 x->pred_mv_sad[i] = INT_MAX;
3203 *returnrate = INT_MAX;
3205 assert(vp9_segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP));
3207 mbmi->mode = ZEROMV;
3208 mbmi->uv_mode = DC_PRED;
3209 mbmi->ref_frame[0] = LAST_FRAME;
3210 mbmi->ref_frame[1] = NONE;
3211 mbmi->mv[0].as_int = 0;
3214 // Search for best switchable filter by checking the variance of
3215 // pred error irrespective of whether the filter will be used
3216 rd_opt->mask_filter = 0;
3217 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
3218 rd_opt->filter_cache[i] = INT64_MAX;
3220 if (cm->interp_filter != BILINEAR) {
3221 best_filter = EIGHTTAP;
3222 if (cm->interp_filter == SWITCHABLE &&
3223 x->source_variance >= cpi->sf.disable_filter_search_var_thresh) {
3225 int best_rs = INT_MAX;
3226 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
3227 mbmi->interp_filter = i;
3228 rs = vp9_get_switchable_rate(cpi);
3231 best_filter = mbmi->interp_filter;
3236 // Set the appropriate filter
3237 if (cm->interp_filter == SWITCHABLE) {
3238 mbmi->interp_filter = best_filter;
3239 rate2 += vp9_get_switchable_rate(cpi);
3241 mbmi->interp_filter = cm->interp_filter;
3244 if (cm->reference_mode == REFERENCE_MODE_SELECT)
3245 rate2 += vp9_cost_bit(comp_mode_p, comp_pred);
3247 // Estimate the reference frame signaling cost and add it
3248 // to the rolling cost variable.
3249 rate2 += ref_costs_single[LAST_FRAME];
3250 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3252 *returnrate = rate2;
3253 *returndistortion = distortion2;
3255 if (this_rd >= best_rd_so_far)
3258 assert((cm->interp_filter == SWITCHABLE) ||
3259 (cm->interp_filter == mbmi->interp_filter));
3261 update_rd_thresh_fact(cpi, bsize, THR_ZEROMV);
3263 vp9_zero(best_pred_diff);
3264 vp9_zero(best_filter_diff);
3265 vp9_zero(best_tx_diff);
3267 if (!x->select_tx_size)
3268 swap_block_ptr(x, ctx, 1, 0, 0, MAX_MB_PLANE);
3269 store_coding_context(x, ctx, THR_ZEROMV,
3270 best_pred_diff, best_tx_diff, best_filter_diff, 0);
3275 int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
3276 const TileInfo *const tile,
3277 int mi_row, int mi_col,
3279 int64_t *returndistortion,
3281 PICK_MODE_CONTEXT *ctx,
3282 int64_t best_rd_so_far) {
3283 VP9_COMMON *const cm = &cpi->common;
3284 RD_OPT *const rd_opt = &cpi->rd;
3285 MACROBLOCKD *const xd = &x->e_mbd;
3286 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
3287 const struct segmentation *const seg = &cm->seg;
3288 MV_REFERENCE_FRAME ref_frame, second_ref_frame;
3289 unsigned char segment_id = mbmi->segment_id;
3291 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
3292 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
3293 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
3295 int64_t best_rd = best_rd_so_far;
3296 int64_t best_yrd = best_rd_so_far; // FIXME(rbultje) more precise
3297 static const int64_t best_tx_diff[TX_MODES] = { 0 };
3298 int64_t best_pred_diff[REFERENCE_MODES];
3299 int64_t best_pred_rd[REFERENCE_MODES];
3300 int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
3301 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3302 MB_MODE_INFO best_mbmode;
3303 int ref_index, best_ref_index = 0;
3304 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3305 vp9_prob comp_mode_p;
3306 INTERP_FILTER tmp_best_filter = SWITCHABLE;
3307 int rate_uv_intra, rate_uv_tokenonly;
3310 PREDICTION_MODE mode_uv = DC_PRED;
3311 int intra_cost_penalty = 20 * vp9_dc_quant(cm->base_qindex, cm->y_dc_delta_q);
3312 int_mv seg_mvs[4][MAX_REF_FRAMES];
3313 b_mode_info best_bmodes[4];
3315 int ref_frame_skip_mask[2] = { 0 };
3317 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3318 vpx_memset(x->zcoeff_blk[TX_4X4], 0, 4);
3319 vp9_zero(best_mbmode);
3321 for (i = 0; i < 4; i++) {
3323 for (j = 0; j < MAX_REF_FRAMES; j++)
3324 seg_mvs[i][j].as_int = INVALID_MV;
3327 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3330 for (i = 0; i < REFERENCE_MODES; ++i)
3331 best_pred_rd[i] = INT64_MAX;
3332 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3333 best_filter_rd[i] = INT64_MAX;
3334 rate_uv_intra = INT_MAX;
3336 *returnrate = INT_MAX;
3338 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ref_frame++) {
3339 if (cpi->ref_frame_flags & flag_list[ref_frame]) {
3340 setup_buffer_inter(cpi, x, tile,
3341 ref_frame, bsize, mi_row, mi_col,
3342 frame_mv[NEARESTMV], frame_mv[NEARMV],
3345 ref_frame_skip_mask[0] |= (1 << ref_frame);
3346 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3348 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
3349 frame_mv[ZEROMV][ref_frame].as_int = 0;
3352 for (ref_index = 0; ref_index < MAX_REFS; ++ref_index) {
3353 int mode_excluded = 0;
3354 int64_t this_rd = INT64_MAX;
3355 int disable_skip = 0;
3356 int compmode_cost = 0;
3357 int rate2 = 0, rate_y = 0, rate_uv = 0;
3358 int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
3362 int64_t total_sse = INT_MAX;
3365 ref_frame = vp9_ref_order[ref_index].ref_frame[0];
3366 second_ref_frame = vp9_ref_order[ref_index].ref_frame[1];
3368 // Look at the reference frame of the best mode so far and set the
3369 // skip mask to look at a subset of the remaining modes.
3370 if (ref_index > 2 && cpi->sf.mode_skip_start < MAX_MODES) {
3371 if (ref_index == 3) {
3372 switch (vp9_ref_order[best_ref_index].ref_frame[0]) {
3376 ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME);
3377 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3380 ref_frame_skip_mask[0] |= (1 << LAST_FRAME) | (1 << ALTREF_FRAME);
3381 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3384 ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << LAST_FRAME);
3387 case MAX_REF_FRAMES:
3388 assert(0 && "Invalid Reference frame");
3394 if (ref_frame_skip_mask[0] & (1 << ref_index) &&
3395 ref_frame_skip_mask[1] & (1 << ref_index))
3398 // Test best rd so far against threshold for trying this mode.
3399 if (rd_less_than_thresh(best_rd,
3400 rd_opt->threshes[segment_id][bsize][ref_index],
3401 rd_opt->thresh_freq_fact[bsize][ref_index]))
3404 if (ref_frame > INTRA_FRAME &&
3405 !(cpi->ref_frame_flags & flag_list[ref_frame])) {
3409 comp_pred = second_ref_frame > INTRA_FRAME;
3411 if (!cm->allow_comp_inter_inter)
3414 if (!(cpi->ref_frame_flags & flag_list[second_ref_frame]))
3416 // Do not allow compound prediction if the segment level reference frame
3417 // feature is in use as in this case there can only be one reference.
3418 if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
3420 if ((cpi->sf.mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
3421 vp9_ref_order[best_ref_index].ref_frame[0] == INTRA_FRAME)
3425 // TODO(jingning, jkoleszar): scaling reference frame not supported for
3427 if (ref_frame > INTRA_FRAME &&
3428 vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
3431 if (second_ref_frame > INTRA_FRAME &&
3432 vp9_is_scaled(&cm->frame_refs[second_ref_frame - 1].sf))
3436 mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
3437 else if (ref_frame != INTRA_FRAME)
3438 mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
3440 // If the segment reference frame feature is enabled....
3441 // then do nothing if the current ref frame is not allowed..
3442 if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
3443 vp9_get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
3445 // Disable this drop out case if the ref frame
3446 // segment level feature is enabled for this segment. This is to
3447 // prevent the possibility that we end up unable to pick any mode.
3448 } else if (!vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
3449 // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
3450 // unless ARNR filtering is enabled in which case we want
3451 // an unfiltered alternative. We allow near/nearest as well
3452 // because they may result in zero-zero MVs but be cheaper.
3453 if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0))
3457 mbmi->tx_size = TX_4X4;
3458 mbmi->uv_mode = DC_PRED;
3459 mbmi->ref_frame[0] = ref_frame;
3460 mbmi->ref_frame[1] = second_ref_frame;
3461 // Evaluate all sub-pel filters irrespective of whether we can use
3462 // them for this frame.
3463 mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
3464 : cm->interp_filter;
3466 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
3468 // Select prediction reference frames.
3469 for (i = 0; i < MAX_MB_PLANE; i++) {
3470 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
3472 xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
3475 if (ref_frame == INTRA_FRAME) {
3477 if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate, &rate_y,
3478 &distortion_y, best_rd) >= best_rd)
3481 rate2 += intra_cost_penalty;
3482 distortion2 += distortion_y;
3484 if (rate_uv_intra == INT_MAX) {
3485 choose_intra_uv_mode(cpi, ctx, bsize, TX_4X4,
3491 rate2 += rate_uv_intra;
3492 rate_uv = rate_uv_tokenonly;
3493 distortion2 += dist_uv;
3494 distortion_uv = dist_uv;
3495 mbmi->uv_mode = mode_uv;
3499 int64_t this_rd_thresh;
3500 int64_t tmp_rd, tmp_best_rd = INT64_MAX, tmp_best_rdu = INT64_MAX;
3501 int tmp_best_rate = INT_MAX, tmp_best_ratey = INT_MAX;
3502 int64_t tmp_best_distortion = INT_MAX, tmp_best_sse, uv_sse;
3503 int tmp_best_skippable = 0;
3504 int switchable_filter_index;
3505 int_mv *second_ref = comp_pred ?
3506 &mbmi->ref_mvs[second_ref_frame][0] : NULL;
3507 b_mode_info tmp_best_bmodes[16];
3508 MB_MODE_INFO tmp_best_mbmode;
3509 BEST_SEG_INFO bsi[SWITCHABLE_FILTERS];
3510 int pred_exists = 0;
3513 this_rd_thresh = (ref_frame == LAST_FRAME) ?
3514 rd_opt->threshes[segment_id][bsize][THR_LAST] :
3515 rd_opt->threshes[segment_id][bsize][THR_ALTR];
3516 this_rd_thresh = (ref_frame == GOLDEN_FRAME) ?
3517 rd_opt->threshes[segment_id][bsize][THR_GOLD] : this_rd_thresh;
3518 rd_opt->mask_filter = 0;
3519 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
3520 rd_opt->filter_cache[i] = INT64_MAX;
3522 if (cm->interp_filter != BILINEAR) {
3523 tmp_best_filter = EIGHTTAP;
3524 if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) {
3525 tmp_best_filter = EIGHTTAP;
3526 } else if (cpi->sf.adaptive_pred_interp_filter == 1 &&
3527 ctx->pred_interp_filter < SWITCHABLE) {
3528 tmp_best_filter = ctx->pred_interp_filter;
3529 } else if (cpi->sf.adaptive_pred_interp_filter == 2) {
3530 tmp_best_filter = ctx->pred_interp_filter < SWITCHABLE ?
3531 ctx->pred_interp_filter : 0;
3533 for (switchable_filter_index = 0;
3534 switchable_filter_index < SWITCHABLE_FILTERS;
3535 ++switchable_filter_index) {
3538 mbmi->interp_filter = switchable_filter_index;
3539 tmp_rd = rd_pick_best_sub8x8_mode(cpi, x, tile,
3540 &mbmi->ref_mvs[ref_frame][0],
3541 second_ref, best_yrd, &rate,
3542 &rate_y, &distortion,
3543 &skippable, &total_sse,
3544 (int) this_rd_thresh, seg_mvs,
3545 bsi, switchable_filter_index,
3548 if (tmp_rd == INT64_MAX)
3550 rs = vp9_get_switchable_rate(cpi);
3551 rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
3552 rd_opt->filter_cache[switchable_filter_index] = tmp_rd;
3553 rd_opt->filter_cache[SWITCHABLE_FILTERS] =
3554 MIN(rd_opt->filter_cache[SWITCHABLE_FILTERS],
3556 if (cm->interp_filter == SWITCHABLE)
3559 rd_opt->mask_filter = MAX(rd_opt->mask_filter, tmp_rd);
3561 newbest = (tmp_rd < tmp_best_rd);
3563 tmp_best_filter = mbmi->interp_filter;
3564 tmp_best_rd = tmp_rd;
3566 if ((newbest && cm->interp_filter == SWITCHABLE) ||
3567 (mbmi->interp_filter == cm->interp_filter &&
3568 cm->interp_filter != SWITCHABLE)) {
3569 tmp_best_rdu = tmp_rd;
3570 tmp_best_rate = rate;
3571 tmp_best_ratey = rate_y;
3572 tmp_best_distortion = distortion;
3573 tmp_best_sse = total_sse;
3574 tmp_best_skippable = skippable;
3575 tmp_best_mbmode = *mbmi;
3576 for (i = 0; i < 4; i++) {
3577 tmp_best_bmodes[i] = xd->mi[0]->bmi[i];
3578 x->zcoeff_blk[TX_4X4][i] = !x->plane[0].eobs[i];
3581 if (switchable_filter_index == 0 &&
3582 cpi->sf.use_rd_breakout &&
3583 best_rd < INT64_MAX) {
3584 if (tmp_best_rdu / 2 > best_rd) {
3585 // skip searching the other filters if the first is
3586 // already substantially larger than the best so far
3587 tmp_best_filter = mbmi->interp_filter;
3588 tmp_best_rdu = INT64_MAX;
3593 } // switchable_filter_index loop
3597 if (tmp_best_rdu == INT64_MAX && pred_exists)
3600 mbmi->interp_filter = (cm->interp_filter == SWITCHABLE ?
3601 tmp_best_filter : cm->interp_filter);
3603 // Handles the special case when a filter that is not in the
3604 // switchable list (bilinear, 6-tap) is indicated at the frame level
3605 tmp_rd = rd_pick_best_sub8x8_mode(cpi, x, tile,
3606 &mbmi->ref_mvs[ref_frame][0],
3607 second_ref, best_yrd, &rate, &rate_y,
3608 &distortion, &skippable, &total_sse,
3609 (int) this_rd_thresh, seg_mvs, bsi, 0,
3611 if (tmp_rd == INT64_MAX)
3614 total_sse = tmp_best_sse;
3615 rate = tmp_best_rate;
3616 rate_y = tmp_best_ratey;
3617 distortion = tmp_best_distortion;
3618 skippable = tmp_best_skippable;
3619 *mbmi = tmp_best_mbmode;
3620 for (i = 0; i < 4; i++)
3621 xd->mi[0]->bmi[i] = tmp_best_bmodes[i];
3625 distortion2 += distortion;
3627 if (cm->interp_filter == SWITCHABLE)
3628 rate2 += vp9_get_switchable_rate(cpi);
3631 mode_excluded = comp_pred ? cm->reference_mode == SINGLE_REFERENCE
3632 : cm->reference_mode == COMPOUND_REFERENCE;
3634 compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
3636 tmp_best_rdu = best_rd -
3637 MIN(RDCOST(x->rdmult, x->rddiv, rate2, distortion2),
3638 RDCOST(x->rdmult, x->rddiv, 0, total_sse));
3640 if (tmp_best_rdu > 0) {
3641 // If even the 'Y' rd value of split is higher than best so far
3642 // then dont bother looking at UV
3643 vp9_build_inter_predictors_sbuv(&x->e_mbd, mi_row, mi_col,
3645 super_block_uvrd(cpi, x, &rate_uv, &distortion_uv, &uv_skippable,
3646 &uv_sse, BLOCK_8X8, tmp_best_rdu);
3647 if (rate_uv == INT_MAX)
3650 distortion2 += distortion_uv;
3651 skippable = skippable && uv_skippable;
3652 total_sse += uv_sse;
3656 if (cm->reference_mode == REFERENCE_MODE_SELECT)
3657 rate2 += compmode_cost;
3659 // Estimate the reference frame signaling cost and add it
3660 // to the rolling cost variable.
3661 if (second_ref_frame > INTRA_FRAME) {
3662 rate2 += ref_costs_comp[ref_frame];
3664 rate2 += ref_costs_single[ref_frame];
3667 if (!disable_skip) {
3668 // Skip is never coded at the segment level for sub8x8 blocks and instead
3669 // always coded in the bitstream at the mode info level.
3671 if (ref_frame != INTRA_FRAME && !xd->lossless) {
3672 if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
3673 RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
3674 // Add in the cost of the no skip flag.
3675 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
3677 // FIXME(rbultje) make this work for splitmv also
3678 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
3679 distortion2 = total_sse;
3680 assert(total_sse >= 0);
3681 rate2 -= (rate_y + rate_uv);
3687 // Add in the cost of the no skip flag.
3688 rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
3691 // Calculate the final RD estimate for this mode.
3692 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3695 if (!disable_skip && ref_frame == INTRA_FRAME) {
3696 for (i = 0; i < REFERENCE_MODES; ++i)
3697 best_pred_rd[i] = MIN(best_pred_rd[i], this_rd);
3698 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3699 best_filter_rd[i] = MIN(best_filter_rd[i], this_rd);
3702 // Did this mode help.. i.e. is it the new best mode
3703 if (this_rd < best_rd || x->skip) {
3704 if (!mode_excluded) {
3705 int max_plane = MAX_MB_PLANE;
3706 // Note index of best mode so far
3707 best_ref_index = ref_index;
3709 if (ref_frame == INTRA_FRAME) {
3710 /* required for left and above block mv */
3711 mbmi->mv[0].as_int = 0;
3715 *returnrate = rate2;
3716 *returndistortion = distortion2;
3718 best_yrd = best_rd -
3719 RDCOST(x->rdmult, x->rddiv, rate_uv, distortion_uv);
3720 best_mbmode = *mbmi;
3721 best_skip2 = this_skip2;
3722 if (!x->select_tx_size)
3723 swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
3724 vpx_memcpy(ctx->zcoeff_blk, x->zcoeff_blk[TX_4X4],
3725 sizeof(uint8_t) * ctx->num_4x4_blk);
3727 for (i = 0; i < 4; i++)
3728 best_bmodes[i] = xd->mi[0]->bmi[i];
3730 // TODO(debargha): enhance this test with a better distortion prediction
3731 // based on qp, activity mask and history
3732 if ((cpi->sf.mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
3733 (ref_index > MIN_EARLY_TERM_INDEX)) {
3734 const int qstep = xd->plane[0].dequant[1];
3735 // TODO(debargha): Enhance this by specializing for each mode_index
3737 if (x->source_variance < UINT_MAX) {
3738 const int var_adjust = (x->source_variance < 16);
3739 scale -= var_adjust;
3741 if (ref_frame > INTRA_FRAME &&
3742 distortion2 * scale < qstep * qstep) {
3749 /* keep record of best compound/single-only prediction */
3750 if (!disable_skip && ref_frame != INTRA_FRAME) {
3751 int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
3753 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
3754 single_rate = rate2 - compmode_cost;
3755 hybrid_rate = rate2;
3757 single_rate = rate2;
3758 hybrid_rate = rate2 + compmode_cost;
3761 single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
3762 hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
3764 if (!comp_pred && single_rd < best_pred_rd[SINGLE_REFERENCE])
3765 best_pred_rd[SINGLE_REFERENCE] = single_rd;
3766 else if (comp_pred && single_rd < best_pred_rd[COMPOUND_REFERENCE])
3767 best_pred_rd[COMPOUND_REFERENCE] = single_rd;
3769 if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
3770 best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
3773 /* keep record of best filter type */
3774 if (!mode_excluded && !disable_skip && ref_frame != INTRA_FRAME &&
3775 cm->interp_filter != BILINEAR) {
3776 int64_t ref = rd_opt->filter_cache[cm->interp_filter == SWITCHABLE ?
3777 SWITCHABLE_FILTERS : cm->interp_filter];
3779 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3780 if (ref == INT64_MAX)
3782 else if (rd_opt->filter_cache[i] == INT64_MAX)
3783 // when early termination is triggered, the encoder does not have
3784 // access to the rate-distortion cost. it only knows that the cost
3785 // should be above the maximum valid value. hence it takes the known
3786 // maximum plus an arbitrary constant as the rate-distortion cost.
3787 adj_rd = rd_opt->mask_filter - ref + 10;
3789 adj_rd = rd_opt->filter_cache[i] - ref;
3792 best_filter_rd[i] = MIN(best_filter_rd[i], adj_rd);
3799 if (x->skip && !comp_pred)
3803 if (best_rd >= best_rd_so_far)
3806 // If we used an estimate for the uv intra rd in the loop above...
3807 if (cpi->sf.use_uv_intra_rd_estimate) {
3808 // Do Intra UV best rd mode selection if best mode choice above was intra.
3809 if (vp9_ref_order[best_ref_index].ref_frame[0] == INTRA_FRAME) {
3810 *mbmi = best_mbmode;
3811 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra,
3819 if (best_rd == INT64_MAX) {
3820 *returnrate = INT_MAX;
3821 *returndistortion = INT64_MAX;
3825 assert((cm->interp_filter == SWITCHABLE) ||
3826 (cm->interp_filter == best_mbmode.interp_filter) ||
3827 !is_inter_block(&best_mbmode));
3829 update_rd_thresh_fact(cpi, bsize, best_ref_index);
3832 *mbmi = best_mbmode;
3833 x->skip |= best_skip2;
3834 if (!is_inter_block(&best_mbmode)) {
3835 for (i = 0; i < 4; i++)
3836 xd->mi[0]->bmi[i].as_mode = best_bmodes[i].as_mode;
3838 for (i = 0; i < 4; ++i)
3839 vpx_memcpy(&xd->mi[0]->bmi[i], &best_bmodes[i], sizeof(b_mode_info));
3841 mbmi->mv[0].as_int = xd->mi[0]->bmi[3].as_mv[0].as_int;
3842 mbmi->mv[1].as_int = xd->mi[0]->bmi[3].as_mv[1].as_int;
3845 for (i = 0; i < REFERENCE_MODES; ++i) {
3846 if (best_pred_rd[i] == INT64_MAX)
3847 best_pred_diff[i] = INT_MIN;
3849 best_pred_diff[i] = best_rd - best_pred_rd[i];
3853 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3854 if (best_filter_rd[i] == INT64_MAX)
3855 best_filter_diff[i] = 0;
3857 best_filter_diff[i] = best_rd - best_filter_rd[i];
3859 if (cm->interp_filter == SWITCHABLE)
3860 assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
3862 vp9_zero(best_filter_diff);
3865 store_coding_context(x, ctx, best_ref_index,
3866 best_pred_diff, best_tx_diff, best_filter_diff, 0);