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"
15 #include "./vpx_dsp_rtcd.h"
17 #include "vpx_dsp/vpx_dsp_common.h"
18 #include "vpx_mem/vpx_mem.h"
19 #include "vpx_ports/mem.h"
20 #include "vpx_ports/system_state.h"
22 #include "vp9/common/vp9_common.h"
23 #include "vp9/common/vp9_entropy.h"
24 #include "vp9/common/vp9_entropymode.h"
25 #include "vp9/common/vp9_idct.h"
26 #include "vp9/common/vp9_mvref_common.h"
27 #include "vp9/common/vp9_pred_common.h"
28 #include "vp9/common/vp9_quant_common.h"
29 #include "vp9/common/vp9_reconinter.h"
30 #include "vp9/common/vp9_reconintra.h"
31 #include "vp9/common/vp9_scan.h"
32 #include "vp9/common/vp9_seg_common.h"
34 #include "vp9/encoder/vp9_cost.h"
35 #include "vp9/encoder/vp9_encodemb.h"
36 #include "vp9/encoder/vp9_encodemv.h"
37 #include "vp9/encoder/vp9_encoder.h"
38 #include "vp9/encoder/vp9_mcomp.h"
39 #include "vp9/encoder/vp9_quantize.h"
40 #include "vp9/encoder/vp9_ratectrl.h"
41 #include "vp9/encoder/vp9_rd.h"
42 #include "vp9/encoder/vp9_rdopt.h"
43 #include "vp9/encoder/vp9_aq_variance.h"
45 #define LAST_FRAME_MODE_MASK \
46 ((1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME) | (1 << INTRA_FRAME))
47 #define GOLDEN_FRAME_MODE_MASK \
48 ((1 << LAST_FRAME) | (1 << ALTREF_FRAME) | (1 << INTRA_FRAME))
49 #define ALT_REF_MODE_MASK \
50 ((1 << LAST_FRAME) | (1 << GOLDEN_FRAME) | (1 << INTRA_FRAME))
52 #define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | 0x01)
54 #define MIN_EARLY_TERM_INDEX 3
55 #define NEW_MV_DISCOUNT_FACTOR 8
59 MV_REFERENCE_FRAME ref_frame[2];
63 MV_REFERENCE_FRAME ref_frame[2];
66 struct rdcost_block_args {
69 ENTROPY_CONTEXT t_above[16];
70 ENTROPY_CONTEXT t_left[16];
77 int use_fast_coef_costing;
82 #define LAST_NEW_MV_INDEX 6
83 static const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
84 { NEARESTMV, { LAST_FRAME, NONE } },
85 { NEARESTMV, { ALTREF_FRAME, NONE } },
86 { NEARESTMV, { GOLDEN_FRAME, NONE } },
88 { DC_PRED, { INTRA_FRAME, NONE } },
90 { NEWMV, { LAST_FRAME, NONE } },
91 { NEWMV, { ALTREF_FRAME, NONE } },
92 { NEWMV, { GOLDEN_FRAME, NONE } },
94 { NEARMV, { LAST_FRAME, NONE } },
95 { NEARMV, { ALTREF_FRAME, NONE } },
96 { NEARMV, { GOLDEN_FRAME, NONE } },
98 { ZEROMV, { LAST_FRAME, NONE } },
99 { ZEROMV, { GOLDEN_FRAME, NONE } },
100 { ZEROMV, { ALTREF_FRAME, NONE } },
102 { NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
103 { NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
105 { TM_PRED, { INTRA_FRAME, NONE } },
107 { NEARMV, { LAST_FRAME, ALTREF_FRAME } },
108 { NEWMV, { LAST_FRAME, ALTREF_FRAME } },
109 { NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
110 { NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
112 { ZEROMV, { LAST_FRAME, ALTREF_FRAME } },
113 { ZEROMV, { GOLDEN_FRAME, ALTREF_FRAME } },
115 { H_PRED, { INTRA_FRAME, NONE } },
116 { V_PRED, { INTRA_FRAME, NONE } },
117 { D135_PRED, { INTRA_FRAME, NONE } },
118 { D207_PRED, { INTRA_FRAME, NONE } },
119 { D153_PRED, { INTRA_FRAME, NONE } },
120 { D63_PRED, { INTRA_FRAME, NONE } },
121 { D117_PRED, { INTRA_FRAME, NONE } },
122 { D45_PRED, { INTRA_FRAME, NONE } },
125 static const REF_DEFINITION vp9_ref_order[MAX_REFS] = {
126 { { LAST_FRAME, NONE } }, { { GOLDEN_FRAME, NONE } },
127 { { ALTREF_FRAME, NONE } }, { { LAST_FRAME, ALTREF_FRAME } },
128 { { GOLDEN_FRAME, ALTREF_FRAME } }, { { INTRA_FRAME, NONE } },
131 static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx, int m, int n,
132 int min_plane, int max_plane) {
135 for (i = min_plane; i < max_plane; ++i) {
136 struct macroblock_plane *const p = &x->plane[i];
137 struct macroblockd_plane *const pd = &x->e_mbd.plane[i];
139 p->coeff = ctx->coeff_pbuf[i][m];
140 p->qcoeff = ctx->qcoeff_pbuf[i][m];
141 pd->dqcoeff = ctx->dqcoeff_pbuf[i][m];
142 p->eobs = ctx->eobs_pbuf[i][m];
144 ctx->coeff_pbuf[i][m] = ctx->coeff_pbuf[i][n];
145 ctx->qcoeff_pbuf[i][m] = ctx->qcoeff_pbuf[i][n];
146 ctx->dqcoeff_pbuf[i][m] = ctx->dqcoeff_pbuf[i][n];
147 ctx->eobs_pbuf[i][m] = ctx->eobs_pbuf[i][n];
149 ctx->coeff_pbuf[i][n] = p->coeff;
150 ctx->qcoeff_pbuf[i][n] = p->qcoeff;
151 ctx->dqcoeff_pbuf[i][n] = pd->dqcoeff;
152 ctx->eobs_pbuf[i][n] = p->eobs;
156 static void model_rd_for_sb(VP9_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x,
157 MACROBLOCKD *xd, int *out_rate_sum,
158 int64_t *out_dist_sum, int *skip_txfm_sb,
159 int64_t *skip_sse_sb) {
160 // Note our transform coeffs are 8 times an orthogonal transform.
161 // Hence quantizer step is also 8 times. To get effective quantizer
162 // we need to divide by 8 before sending to modeling function.
164 int64_t rate_sum = 0;
165 int64_t dist_sum = 0;
166 const int ref = xd->mi[0]->ref_frame[0];
168 unsigned int var = 0;
169 int64_t total_sse = 0;
173 const int dequant_shift =
174 #if CONFIG_VP9_HIGHBITDEPTH
175 (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd - 5 :
176 #endif // CONFIG_VP9_HIGHBITDEPTH
178 unsigned int qstep_vec[MAX_MB_PLANE];
179 unsigned int nlog2_vec[MAX_MB_PLANE];
180 unsigned int sum_sse_vec[MAX_MB_PLANE];
181 int any_zero_sum_sse = 0;
183 x->pred_sse[ref] = 0;
185 for (i = 0; i < MAX_MB_PLANE; ++i) {
186 struct macroblock_plane *const p = &x->plane[i];
187 struct macroblockd_plane *const pd = &xd->plane[i];
188 const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
189 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
190 const BLOCK_SIZE unit_size = txsize_to_bsize[max_tx_size];
191 const int64_t dc_thr = p->quant_thred[0] >> shift;
192 const int64_t ac_thr = p->quant_thred[1] >> shift;
193 unsigned int sum_sse = 0;
194 // The low thresholds are used to measure if the prediction errors are
195 // low enough so that we can skip the mode search.
196 const int64_t low_dc_thr = VPXMIN(50, dc_thr >> 2);
197 const int64_t low_ac_thr = VPXMIN(80, ac_thr >> 2);
198 int bw = 1 << (b_width_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
199 int bh = 1 << (b_height_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
201 int lw = b_width_log2_lookup[unit_size] + 2;
202 int lh = b_height_log2_lookup[unit_size] + 2;
204 for (idy = 0; idy < bh; ++idy) {
205 for (idx = 0; idx < bw; ++idx) {
206 uint8_t *src = p->src.buf + (idy * p->src.stride << lh) + (idx << lw);
207 uint8_t *dst = pd->dst.buf + (idy * pd->dst.stride << lh) + (idx << lh);
208 int block_idx = (idy << 1) + idx;
209 int low_err_skip = 0;
211 var = cpi->fn_ptr[unit_size].vf(src, p->src.stride, dst, pd->dst.stride,
213 x->bsse[(i << 2) + block_idx] = sse;
216 x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_NONE;
217 if (!x->select_tx_size) {
218 // Check if all ac coefficients can be quantized to zero.
219 if (var < ac_thr || var == 0) {
220 x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_ONLY;
222 // Check if dc coefficient can be quantized to zero.
223 if (sse - var < dc_thr || sse == var) {
224 x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_DC;
226 if (!sse || (var < low_ac_thr && sse - var < low_dc_thr))
232 if (skip_flag && !low_err_skip) skip_flag = 0;
234 if (i == 0) x->pred_sse[ref] += sse;
238 total_sse += sum_sse;
239 sum_sse_vec[i] = sum_sse;
240 any_zero_sum_sse = any_zero_sum_sse || (sum_sse == 0);
241 qstep_vec[i] = pd->dequant[1] >> dequant_shift;
242 nlog2_vec[i] = num_pels_log2_lookup[bs];
245 // Fast approximate the modelling function.
246 if (cpi->sf.simple_model_rd_from_var) {
247 for (i = 0; i < MAX_MB_PLANE; ++i) {
249 const int64_t square_error = sum_sse_vec[i];
250 int quantizer = qstep_vec[i];
253 rate = (square_error * (280 - quantizer)) >> (16 - VP9_PROB_COST_SHIFT);
256 dist = (square_error * quantizer) >> 8;
261 if (any_zero_sum_sse) {
262 for (i = 0; i < MAX_MB_PLANE; ++i) {
264 vp9_model_rd_from_var_lapndz(sum_sse_vec[i], nlog2_vec[i], qstep_vec[i],
270 vp9_model_rd_from_var_lapndz_vec(sum_sse_vec, nlog2_vec, qstep_vec,
271 &rate_sum, &dist_sum);
275 *skip_txfm_sb = skip_flag;
276 *skip_sse_sb = total_sse << 4;
277 *out_rate_sum = (int)rate_sum;
278 *out_dist_sum = dist_sum << 4;
281 #if CONFIG_VP9_HIGHBITDEPTH
282 int64_t vp9_highbd_block_error_c(const tran_low_t *coeff,
283 const tran_low_t *dqcoeff, intptr_t block_size,
284 int64_t *ssz, int bd) {
286 int64_t error = 0, sqcoeff = 0;
287 int shift = 2 * (bd - 8);
288 int rounding = shift > 0 ? 1 << (shift - 1) : 0;
290 for (i = 0; i < block_size; i++) {
291 const int64_t diff = coeff[i] - dqcoeff[i];
292 error += diff * diff;
293 sqcoeff += (int64_t)coeff[i] * (int64_t)coeff[i];
295 assert(error >= 0 && sqcoeff >= 0);
296 error = (error + rounding) >> shift;
297 sqcoeff = (sqcoeff + rounding) >> shift;
303 static int64_t vp9_highbd_block_error_dispatch(const tran_low_t *coeff,
304 const tran_low_t *dqcoeff,
306 int64_t *ssz, int bd) {
308 return vp9_block_error(coeff, dqcoeff, block_size, ssz);
310 return vp9_highbd_block_error(coeff, dqcoeff, block_size, ssz, bd);
313 #endif // CONFIG_VP9_HIGHBITDEPTH
315 int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
316 intptr_t block_size, int64_t *ssz) {
318 int64_t error = 0, sqcoeff = 0;
320 for (i = 0; i < block_size; i++) {
321 const int diff = coeff[i] - dqcoeff[i];
322 error += diff * diff;
323 sqcoeff += coeff[i] * coeff[i];
330 int64_t vp9_block_error_fp_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
335 for (i = 0; i < block_size; i++) {
336 const int diff = coeff[i] - dqcoeff[i];
337 error += diff * diff;
343 /* The trailing '0' is a terminator which is used inside cost_coeffs() to
344 * decide whether to include cost of a trailing EOB node or not (i.e. we
345 * can skip this if the last coefficient in this transform block, e.g. the
346 * 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block,
348 static const int16_t band_counts[TX_SIZES][8] = {
349 { 1, 2, 3, 4, 3, 16 - 13, 0 },
350 { 1, 2, 3, 4, 11, 64 - 21, 0 },
351 { 1, 2, 3, 4, 11, 256 - 21, 0 },
352 { 1, 2, 3, 4, 11, 1024 - 21, 0 },
354 static int cost_coeffs(MACROBLOCK *x, int plane, int block, TX_SIZE tx_size,
355 int pt, const int16_t *scan, const int16_t *nb,
356 int use_fast_coef_costing) {
357 MACROBLOCKD *const xd = &x->e_mbd;
358 MODE_INFO *mi = xd->mi[0];
359 const struct macroblock_plane *p = &x->plane[plane];
360 const PLANE_TYPE type = get_plane_type(plane);
361 const int16_t *band_count = &band_counts[tx_size][1];
362 const int eob = p->eobs[block];
363 const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
364 unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
365 x->token_costs[tx_size][type][is_inter_block(mi)];
366 uint8_t token_cache[32 * 32];
368 #if CONFIG_VP9_HIGHBITDEPTH
369 const uint16_t *cat6_high_cost = vp9_get_high_cost_table(xd->bd);
371 const uint16_t *cat6_high_cost = vp9_get_high_cost_table(8);
374 // Check for consistency of tx_size with mode info
375 assert(type == PLANE_TYPE_Y
376 ? mi->tx_size == tx_size
377 : get_uv_tx_size(mi, &xd->plane[plane]) == tx_size);
381 cost = token_costs[0][0][pt][EOB_TOKEN];
383 if (use_fast_coef_costing) {
384 int band_left = *band_count++;
390 cost = vp9_get_token_cost(v, &prev_t, cat6_high_cost);
391 cost += (*token_costs)[0][pt][prev_t];
393 token_cache[0] = vp9_pt_energy_class[prev_t];
397 for (c = 1; c < eob; c++) {
398 const int rc = scan[c];
402 cost += vp9_get_token_cost(v, &t, cat6_high_cost);
403 cost += (*token_costs)[!prev_t][!prev_t][t];
406 band_left = *band_count++;
412 if (band_left) cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
414 } else { // !use_fast_coef_costing
415 int band_left = *band_count++;
421 unsigned int(*tok_cost_ptr)[COEFF_CONTEXTS][ENTROPY_TOKENS];
422 cost = vp9_get_token_cost(v, &tok, cat6_high_cost);
423 cost += (*token_costs)[0][pt][tok];
425 token_cache[0] = vp9_pt_energy_class[tok];
428 tok_cost_ptr = &((*token_costs)[!tok]);
431 for (c = 1; c < eob; c++) {
432 const int rc = scan[c];
435 cost += vp9_get_token_cost(v, &tok, cat6_high_cost);
436 pt = get_coef_context(nb, token_cache, c);
437 cost += (*tok_cost_ptr)[pt][tok];
438 token_cache[rc] = vp9_pt_energy_class[tok];
440 band_left = *band_count++;
443 tok_cost_ptr = &((*token_costs)[!tok]);
448 pt = get_coef_context(nb, token_cache, c);
449 cost += (*token_costs)[0][pt][EOB_TOKEN];
457 static INLINE int num_4x4_to_edge(int plane_4x4_dim, int mb_to_edge_dim,
458 int subsampling_dim, int blk_dim) {
459 return plane_4x4_dim + (mb_to_edge_dim >> (5 + subsampling_dim)) - blk_dim;
462 // Compute the pixel domain sum square error on all visible 4x4s in the
464 static unsigned pixel_sse(const VP9_COMP *const cpi, const MACROBLOCKD *xd,
465 const struct macroblockd_plane *const pd,
466 const uint8_t *src, const int src_stride,
467 const uint8_t *dst, const int dst_stride, int blk_row,
468 int blk_col, const BLOCK_SIZE plane_bsize,
469 const BLOCK_SIZE tx_bsize) {
470 unsigned int sse = 0;
471 const int plane_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
472 const int plane_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
473 const int tx_4x4_w = num_4x4_blocks_wide_lookup[tx_bsize];
474 const int tx_4x4_h = num_4x4_blocks_high_lookup[tx_bsize];
475 int b4x4s_to_right_edge = num_4x4_to_edge(plane_4x4_w, xd->mb_to_right_edge,
476 pd->subsampling_x, blk_col);
477 int b4x4s_to_bottom_edge = num_4x4_to_edge(plane_4x4_h, xd->mb_to_bottom_edge,
478 pd->subsampling_y, blk_row);
479 if (tx_bsize == BLOCK_4X4 ||
480 (b4x4s_to_right_edge >= tx_4x4_w && b4x4s_to_bottom_edge >= tx_4x4_h)) {
481 cpi->fn_ptr[tx_bsize].vf(src, src_stride, dst, dst_stride, &sse);
483 const vpx_variance_fn_t vf_4x4 = cpi->fn_ptr[BLOCK_4X4].vf;
485 unsigned this_sse = 0;
486 int max_r = VPXMIN(b4x4s_to_bottom_edge, tx_4x4_h);
487 int max_c = VPXMIN(b4x4s_to_right_edge, tx_4x4_w);
489 // if we are in the unrestricted motion border.
490 for (r = 0; r < max_r; ++r) {
491 // Skip visiting the sub blocks that are wholly within the UMV.
492 for (c = 0; c < max_c; ++c) {
493 vf_4x4(src + r * src_stride * 4 + c * 4, src_stride,
494 dst + r * dst_stride * 4 + c * 4, dst_stride, &this_sse);
502 // Compute the squares sum squares on all visible 4x4s in the transform block.
503 static int64_t sum_squares_visible(const MACROBLOCKD *xd,
504 const struct macroblockd_plane *const pd,
505 const int16_t *diff, const int diff_stride,
506 int blk_row, int blk_col,
507 const BLOCK_SIZE plane_bsize,
508 const BLOCK_SIZE tx_bsize) {
510 const int plane_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
511 const int plane_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
512 const int tx_4x4_w = num_4x4_blocks_wide_lookup[tx_bsize];
513 const int tx_4x4_h = num_4x4_blocks_high_lookup[tx_bsize];
514 int b4x4s_to_right_edge = num_4x4_to_edge(plane_4x4_w, xd->mb_to_right_edge,
515 pd->subsampling_x, blk_col);
516 int b4x4s_to_bottom_edge = num_4x4_to_edge(plane_4x4_h, xd->mb_to_bottom_edge,
517 pd->subsampling_y, blk_row);
518 if (tx_bsize == BLOCK_4X4 ||
519 (b4x4s_to_right_edge >= tx_4x4_w && b4x4s_to_bottom_edge >= tx_4x4_h)) {
520 assert(tx_4x4_w == tx_4x4_h);
521 sse = (int64_t)vpx_sum_squares_2d_i16(diff, diff_stride, tx_4x4_w << 2);
524 int max_r = VPXMIN(b4x4s_to_bottom_edge, tx_4x4_h);
525 int max_c = VPXMIN(b4x4s_to_right_edge, tx_4x4_w);
527 // if we are in the unrestricted motion border.
528 for (r = 0; r < max_r; ++r) {
529 // Skip visiting the sub blocks that are wholly within the UMV.
530 for (c = 0; c < max_c; ++c) {
531 sse += (int64_t)vpx_sum_squares_2d_i16(
532 diff + r * diff_stride * 4 + c * 4, diff_stride, 4);
539 static void dist_block(const VP9_COMP *cpi, MACROBLOCK *x, int plane,
540 BLOCK_SIZE plane_bsize, int block, int blk_row,
541 int blk_col, TX_SIZE tx_size, int64_t *out_dist,
543 MACROBLOCKD *const xd = &x->e_mbd;
544 const struct macroblock_plane *const p = &x->plane[plane];
545 const struct macroblockd_plane *const pd = &xd->plane[plane];
547 if (x->block_tx_domain) {
548 const int ss_txfrm_size = tx_size << 1;
550 const int shift = tx_size == TX_32X32 ? 0 : 2;
551 const tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
552 const tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
553 #if CONFIG_VP9_HIGHBITDEPTH
554 const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
555 *out_dist = vp9_highbd_block_error_dispatch(
556 coeff, dqcoeff, 16 << ss_txfrm_size, &this_sse, bd) >>
560 vp9_block_error(coeff, dqcoeff, 16 << ss_txfrm_size, &this_sse) >>
562 #endif // CONFIG_VP9_HIGHBITDEPTH
563 *out_sse = this_sse >> shift;
565 if (x->skip_encode && !is_inter_block(xd->mi[0])) {
566 // TODO(jingning): tune the model to better capture the distortion.
568 (pd->dequant[1] * pd->dequant[1] * (1 << ss_txfrm_size)) >>
569 #if CONFIG_VP9_HIGHBITDEPTH
570 (shift + 2 + (bd - 8) * 2);
573 #endif // CONFIG_VP9_HIGHBITDEPTH
574 *out_dist += (p >> 4);
578 const BLOCK_SIZE tx_bsize = txsize_to_bsize[tx_size];
579 const int bs = 4 * num_4x4_blocks_wide_lookup[tx_bsize];
580 const int src_stride = p->src.stride;
581 const int dst_stride = pd->dst.stride;
582 const int src_idx = 4 * (blk_row * src_stride + blk_col);
583 const int dst_idx = 4 * (blk_row * dst_stride + blk_col);
584 const uint8_t *src = &p->src.buf[src_idx];
585 const uint8_t *dst = &pd->dst.buf[dst_idx];
586 const tran_low_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
587 const uint16_t *eob = &p->eobs[block];
590 tmp = pixel_sse(cpi, xd, pd, src, src_stride, dst, dst_stride, blk_row,
591 blk_col, plane_bsize, tx_bsize);
592 *out_sse = (int64_t)tmp * 16;
595 #if CONFIG_VP9_HIGHBITDEPTH
596 DECLARE_ALIGNED(16, uint16_t, recon16[1024]);
597 uint8_t *recon = (uint8_t *)recon16;
599 DECLARE_ALIGNED(16, uint8_t, recon[1024]);
600 #endif // CONFIG_VP9_HIGHBITDEPTH
602 #if CONFIG_VP9_HIGHBITDEPTH
603 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
604 vpx_highbd_convolve_copy(CONVERT_TO_SHORTPTR(dst), dst_stride, recon16,
605 32, NULL, 0, 0, 0, 0, bs, bs, xd->bd);
607 vp9_highbd_iwht4x4_add(dqcoeff, recon16, 32, *eob, xd->bd);
611 vp9_highbd_idct4x4_add(dqcoeff, recon16, 32, *eob, xd->bd);
614 vp9_highbd_idct8x8_add(dqcoeff, recon16, 32, *eob, xd->bd);
617 vp9_highbd_idct16x16_add(dqcoeff, recon16, 32, *eob, xd->bd);
620 vp9_highbd_idct32x32_add(dqcoeff, recon16, 32, *eob, xd->bd);
622 default: assert(0 && "Invalid transform size");
625 recon = CONVERT_TO_BYTEPTR(recon16);
627 #endif // CONFIG_VP9_HIGHBITDEPTH
628 vpx_convolve_copy(dst, dst_stride, recon, 32, NULL, 0, 0, 0, 0, bs, bs);
630 case TX_32X32: vp9_idct32x32_add(dqcoeff, recon, 32, *eob); break;
631 case TX_16X16: vp9_idct16x16_add(dqcoeff, recon, 32, *eob); break;
632 case TX_8X8: vp9_idct8x8_add(dqcoeff, recon, 32, *eob); break;
634 // this is like vp9_short_idct4x4 but has a special case around
635 // eob<=1, which is significant (not just an optimization) for
636 // the lossless case.
637 x->inv_txfm_add(dqcoeff, recon, 32, *eob);
639 default: assert(0 && "Invalid transform size"); break;
641 #if CONFIG_VP9_HIGHBITDEPTH
643 #endif // CONFIG_VP9_HIGHBITDEPTH
645 tmp = pixel_sse(cpi, xd, pd, src, src_stride, recon, 32, blk_row, blk_col,
646 plane_bsize, tx_bsize);
649 *out_dist = (int64_t)tmp * 16;
653 static int rate_block(int plane, int block, TX_SIZE tx_size, int coeff_ctx,
654 struct rdcost_block_args *args) {
655 return cost_coeffs(args->x, plane, block, tx_size, coeff_ctx, args->so->scan,
656 args->so->neighbors, args->use_fast_coef_costing);
659 static void block_rd_txfm(int plane, int block, int blk_row, int blk_col,
660 BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
661 struct rdcost_block_args *args = arg;
662 MACROBLOCK *const x = args->x;
663 MACROBLOCKD *const xd = &x->e_mbd;
664 MODE_INFO *const mi = xd->mi[0];
665 int64_t rd1, rd2, rd;
669 const int coeff_ctx =
670 combine_entropy_contexts(args->t_left[blk_row], args->t_above[blk_col]);
672 if (args->exit_early) return;
674 if (!is_inter_block(mi)) {
675 struct encode_b_args intra_arg = { x, x->block_qcoeff_opt, args->t_above,
676 args->t_left, &mi->skip };
677 vp9_encode_block_intra(plane, block, blk_row, blk_col, plane_bsize, tx_size,
679 if (x->block_tx_domain) {
680 dist_block(args->cpi, x, plane, plane_bsize, block, blk_row, blk_col,
681 tx_size, &dist, &sse);
683 const BLOCK_SIZE tx_bsize = txsize_to_bsize[tx_size];
684 const struct macroblock_plane *const p = &x->plane[plane];
685 const struct macroblockd_plane *const pd = &xd->plane[plane];
686 const int src_stride = p->src.stride;
687 const int dst_stride = pd->dst.stride;
688 const int diff_stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
689 const uint8_t *src = &p->src.buf[4 * (blk_row * src_stride + blk_col)];
690 const uint8_t *dst = &pd->dst.buf[4 * (blk_row * dst_stride + blk_col)];
691 const int16_t *diff = &p->src_diff[4 * (blk_row * diff_stride + blk_col)];
693 sse = sum_squares_visible(xd, pd, diff, diff_stride, blk_row, blk_col,
694 plane_bsize, tx_bsize);
695 #if CONFIG_VP9_HIGHBITDEPTH
696 if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && (xd->bd > 8))
697 sse = ROUND64_POWER_OF_TWO(sse, (xd->bd - 8) * 2);
698 #endif // CONFIG_VP9_HIGHBITDEPTH
700 tmp = pixel_sse(args->cpi, xd, pd, src, src_stride, dst, dst_stride,
701 blk_row, blk_col, plane_bsize, tx_bsize);
702 dist = (int64_t)tmp * 16;
704 } else if (max_txsize_lookup[plane_bsize] == tx_size) {
705 if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] ==
707 // full forward transform and quantization
708 vp9_xform_quant(x, plane, block, blk_row, blk_col, plane_bsize, tx_size);
709 if (x->block_qcoeff_opt)
710 vp9_optimize_b(x, plane, block, tx_size, coeff_ctx);
711 dist_block(args->cpi, x, plane, plane_bsize, block, blk_row, blk_col,
712 tx_size, &dist, &sse);
713 } else if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] ==
715 // compute DC coefficient
716 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[plane].coeff, block);
717 tran_low_t *const dqcoeff = BLOCK_OFFSET(xd->plane[plane].dqcoeff, block);
718 vp9_xform_quant_dc(x, plane, block, blk_row, blk_col, plane_bsize,
720 sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
722 if (x->plane[plane].eobs[block]) {
723 const int64_t orig_sse = (int64_t)coeff[0] * coeff[0];
724 const int64_t resd_sse = coeff[0] - dqcoeff[0];
725 int64_t dc_correct = orig_sse - resd_sse * resd_sse;
726 #if CONFIG_VP9_HIGHBITDEPTH
727 dc_correct >>= ((xd->bd - 8) * 2);
729 if (tx_size != TX_32X32) dc_correct >>= 2;
731 dist = VPXMAX(0, sse - dc_correct);
735 // skip forward transform. Because this is handled here, the quantization
736 // does not need to do it.
737 x->plane[plane].eobs[block] = 0;
738 sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
742 // full forward transform and quantization
743 vp9_xform_quant(x, plane, block, blk_row, blk_col, plane_bsize, tx_size);
744 if (x->block_qcoeff_opt)
745 vp9_optimize_b(x, plane, block, tx_size, coeff_ctx);
746 dist_block(args->cpi, x, plane, plane_bsize, block, blk_row, blk_col,
747 tx_size, &dist, &sse);
750 rd = RDCOST(x->rdmult, x->rddiv, 0, dist);
751 if (args->this_rd + rd > args->best_rd) {
752 args->exit_early = 1;
756 rate = rate_block(plane, block, tx_size, coeff_ctx, args);
757 args->t_above[blk_col] = (x->plane[plane].eobs[block] > 0) ? 1 : 0;
758 args->t_left[blk_row] = (x->plane[plane].eobs[block] > 0) ? 1 : 0;
759 rd1 = RDCOST(x->rdmult, x->rddiv, rate, dist);
760 rd2 = RDCOST(x->rdmult, x->rddiv, 0, sse);
762 // TODO(jingning): temporarily enabled only for luma component
763 rd = VPXMIN(rd1, rd2);
765 x->zcoeff_blk[tx_size][block] =
766 !x->plane[plane].eobs[block] || (rd1 > rd2 && !xd->lossless);
767 x->sum_y_eobs[tx_size] += x->plane[plane].eobs[block];
770 args->this_rate += rate;
771 args->this_dist += dist;
772 args->this_sse += sse;
775 if (args->this_rd > args->best_rd) {
776 args->exit_early = 1;
780 args->skippable &= !x->plane[plane].eobs[block];
783 static void txfm_rd_in_plane(const VP9_COMP *cpi, MACROBLOCK *x, int *rate,
784 int64_t *distortion, int *skippable, int64_t *sse,
785 int64_t ref_best_rd, int plane, BLOCK_SIZE bsize,
786 TX_SIZE tx_size, int use_fast_coef_casting) {
787 MACROBLOCKD *const xd = &x->e_mbd;
788 const struct macroblockd_plane *const pd = &xd->plane[plane];
789 struct rdcost_block_args args;
793 args.best_rd = ref_best_rd;
794 args.use_fast_coef_costing = use_fast_coef_casting;
797 if (plane == 0) xd->mi[0]->tx_size = tx_size;
799 vp9_get_entropy_contexts(bsize, tx_size, pd, args.t_above, args.t_left);
801 args.so = get_scan(xd, tx_size, get_plane_type(plane), 0);
803 vp9_foreach_transformed_block_in_plane(xd, bsize, plane, block_rd_txfm,
805 if (args.exit_early) {
807 *distortion = INT64_MAX;
811 *distortion = args.this_dist;
812 *rate = args.this_rate;
813 *sse = args.this_sse;
814 *skippable = args.skippable;
818 static void choose_largest_tx_size(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
819 int64_t *distortion, int *skip, int64_t *sse,
820 int64_t ref_best_rd, BLOCK_SIZE bs) {
821 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
822 VP9_COMMON *const cm = &cpi->common;
823 const TX_SIZE largest_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
824 MACROBLOCKD *const xd = &x->e_mbd;
825 MODE_INFO *const mi = xd->mi[0];
827 mi->tx_size = VPXMIN(max_tx_size, largest_tx_size);
829 txfm_rd_in_plane(cpi, x, rate, distortion, skip, sse, ref_best_rd, 0, bs,
830 mi->tx_size, cpi->sf.use_fast_coef_costing);
833 static void choose_tx_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
834 int64_t *distortion, int *skip,
835 int64_t *psse, int64_t ref_best_rd,
837 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
838 VP9_COMMON *const cm = &cpi->common;
839 MACROBLOCKD *const xd = &x->e_mbd;
840 MODE_INFO *const mi = xd->mi[0];
841 vpx_prob skip_prob = vp9_get_skip_prob(cm, xd);
842 int r[TX_SIZES][2], s[TX_SIZES];
843 int64_t d[TX_SIZES], sse[TX_SIZES];
844 int64_t rd[TX_SIZES][2] = { { INT64_MAX, INT64_MAX },
845 { INT64_MAX, INT64_MAX },
846 { INT64_MAX, INT64_MAX },
847 { INT64_MAX, INT64_MAX } };
850 int64_t best_rd = INT64_MAX;
851 TX_SIZE best_tx = max_tx_size;
852 int start_tx, end_tx;
854 const vpx_prob *tx_probs = get_tx_probs2(max_tx_size, xd, &cm->fc->tx_probs);
855 assert(skip_prob > 0);
856 s0 = vp9_cost_bit(skip_prob, 0);
857 s1 = vp9_cost_bit(skip_prob, 1);
859 if (cm->tx_mode == TX_MODE_SELECT) {
860 start_tx = max_tx_size;
863 TX_SIZE chosen_tx_size =
864 VPXMIN(max_tx_size, tx_mode_to_biggest_tx_size[cm->tx_mode]);
865 start_tx = chosen_tx_size;
866 end_tx = chosen_tx_size;
869 for (n = start_tx; n >= end_tx; n--) {
871 for (m = 0; m <= n - (n == (int)max_tx_size); m++) {
873 r_tx_size += vp9_cost_zero(tx_probs[m]);
875 r_tx_size += vp9_cost_one(tx_probs[m]);
877 txfm_rd_in_plane(cpi, x, &r[n][0], &d[n], &s[n], &sse[n], ref_best_rd, 0,
878 bs, n, cpi->sf.use_fast_coef_costing);
880 if (r[n][0] < INT_MAX) {
881 r[n][1] += r_tx_size;
883 if (d[n] == INT64_MAX || r[n][0] == INT_MAX) {
884 rd[n][0] = rd[n][1] = INT64_MAX;
886 if (is_inter_block(mi)) {
887 rd[n][0] = rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1, sse[n]);
888 r[n][1] -= r_tx_size;
890 rd[n][0] = RDCOST(x->rdmult, x->rddiv, s1, sse[n]);
891 rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1 + r_tx_size, sse[n]);
894 rd[n][0] = RDCOST(x->rdmult, x->rddiv, r[n][0] + s0, d[n]);
895 rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]);
898 if (is_inter_block(mi) && !xd->lossless && !s[n] && sse[n] != INT64_MAX) {
899 rd[n][0] = VPXMIN(rd[n][0], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
900 rd[n][1] = VPXMIN(rd[n][1], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
903 // Early termination in transform size search.
904 if (cpi->sf.tx_size_search_breakout &&
905 (rd[n][1] == INT64_MAX ||
906 (n < (int)max_tx_size && rd[n][1] > rd[n + 1][1]) || s[n] == 1))
909 if (rd[n][1] < best_rd) {
914 mi->tx_size = best_tx;
916 *distortion = d[mi->tx_size];
917 *rate = r[mi->tx_size][cm->tx_mode == TX_MODE_SELECT];
918 *skip = s[mi->tx_size];
919 *psse = sse[mi->tx_size];
922 static void super_block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
923 int64_t *distortion, int *skip, int64_t *psse,
924 BLOCK_SIZE bs, int64_t ref_best_rd) {
925 MACROBLOCKD *xd = &x->e_mbd;
927 int64_t *ret_sse = psse ? psse : &sse;
929 assert(bs == xd->mi[0]->sb_type);
931 if (cpi->sf.tx_size_search_method == USE_LARGESTALL || xd->lossless) {
932 choose_largest_tx_size(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
935 choose_tx_size_from_rd(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
940 static int conditional_skipintra(PREDICTION_MODE mode,
941 PREDICTION_MODE best_intra_mode) {
942 if (mode == D117_PRED && best_intra_mode != V_PRED &&
943 best_intra_mode != D135_PRED)
945 if (mode == D63_PRED && best_intra_mode != V_PRED &&
946 best_intra_mode != D45_PRED)
948 if (mode == D207_PRED && best_intra_mode != H_PRED &&
949 best_intra_mode != D45_PRED)
951 if (mode == D153_PRED && best_intra_mode != H_PRED &&
952 best_intra_mode != D135_PRED)
957 static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x, int row,
958 int col, PREDICTION_MODE *best_mode,
959 const int *bmode_costs, ENTROPY_CONTEXT *a,
960 ENTROPY_CONTEXT *l, int *bestrate,
961 int *bestratey, int64_t *bestdistortion,
962 BLOCK_SIZE bsize, int64_t rd_thresh) {
963 PREDICTION_MODE mode;
964 MACROBLOCKD *const xd = &x->e_mbd;
965 int64_t best_rd = rd_thresh;
966 struct macroblock_plane *p = &x->plane[0];
967 struct macroblockd_plane *pd = &xd->plane[0];
968 const int src_stride = p->src.stride;
969 const int dst_stride = pd->dst.stride;
970 const uint8_t *src_init = &p->src.buf[row * 4 * src_stride + col * 4];
971 uint8_t *dst_init = &pd->dst.buf[row * 4 * src_stride + col * 4];
972 ENTROPY_CONTEXT ta[2], tempa[2];
973 ENTROPY_CONTEXT tl[2], templ[2];
974 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
975 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
977 uint8_t best_dst[8 * 8];
978 #if CONFIG_VP9_HIGHBITDEPTH
979 uint16_t best_dst16[8 * 8];
981 memcpy(ta, a, num_4x4_blocks_wide * sizeof(a[0]));
982 memcpy(tl, l, num_4x4_blocks_high * sizeof(l[0]));
984 xd->mi[0]->tx_size = TX_4X4;
986 #if CONFIG_VP9_HIGHBITDEPTH
987 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
988 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
991 int64_t distortion = 0;
992 int rate = bmode_costs[mode];
994 if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode))) continue;
996 // Only do the oblique modes if the best so far is
997 // one of the neighboring directional modes
998 if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
999 if (conditional_skipintra(mode, *best_mode)) continue;
1002 memcpy(tempa, ta, num_4x4_blocks_wide * sizeof(ta[0]));
1003 memcpy(templ, tl, num_4x4_blocks_high * sizeof(tl[0]));
1005 for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
1006 for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
1007 const int block = (row + idy) * 2 + (col + idx);
1008 const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
1009 uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
1010 uint16_t *const dst16 = CONVERT_TO_SHORTPTR(dst);
1011 int16_t *const src_diff =
1012 vp9_raster_block_offset_int16(BLOCK_8X8, block, p->src_diff);
1013 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
1014 xd->mi[0]->bmi[block].as_mode = mode;
1015 vp9_predict_intra_block(xd, 1, TX_4X4, mode,
1016 x->skip_encode ? src : dst,
1017 x->skip_encode ? src_stride : dst_stride, dst,
1018 dst_stride, col + idx, row + idy, 0);
1019 vpx_highbd_subtract_block(4, 4, src_diff, 8, src, src_stride, dst,
1020 dst_stride, xd->bd);
1022 const scan_order *so = &vp9_default_scan_orders[TX_4X4];
1023 const int coeff_ctx =
1024 combine_entropy_contexts(tempa[idx], templ[idy]);
1025 vp9_highbd_fwht4x4(src_diff, coeff, 8);
1026 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
1027 ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
1028 so->neighbors, cpi->sf.use_fast_coef_costing);
1029 tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0 ? 1 : 0);
1030 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
1032 vp9_highbd_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block), dst16,
1033 dst_stride, p->eobs[block], xd->bd);
1036 const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
1037 const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
1038 const int coeff_ctx =
1039 combine_entropy_contexts(tempa[idx], templ[idy]);
1040 if (tx_type == DCT_DCT)
1041 vpx_highbd_fdct4x4(src_diff, coeff, 8);
1043 vp9_highbd_fht4x4(src_diff, coeff, 8, tx_type);
1044 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
1045 ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
1046 so->neighbors, cpi->sf.use_fast_coef_costing);
1047 distortion += vp9_highbd_block_error_dispatch(
1048 coeff, BLOCK_OFFSET(pd->dqcoeff, block), 16,
1051 tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0 ? 1 : 0);
1052 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
1054 vp9_highbd_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block),
1055 dst16, dst_stride, p->eobs[block], xd->bd);
1061 this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
1063 if (this_rd < best_rd) {
1066 *bestdistortion = distortion;
1069 memcpy(a, tempa, num_4x4_blocks_wide * sizeof(tempa[0]));
1070 memcpy(l, templ, num_4x4_blocks_high * sizeof(templ[0]));
1071 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
1072 memcpy(best_dst16 + idy * 8,
1073 CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
1074 num_4x4_blocks_wide * 4 * sizeof(uint16_t));
1079 if (best_rd >= rd_thresh || x->skip_encode) return best_rd;
1081 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
1082 memcpy(CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
1083 best_dst16 + idy * 8, num_4x4_blocks_wide * 4 * sizeof(uint16_t));
1088 #endif // CONFIG_VP9_HIGHBITDEPTH
1090 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
1093 int64_t distortion = 0;
1094 int rate = bmode_costs[mode];
1096 if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode))) continue;
1098 // Only do the oblique modes if the best so far is
1099 // one of the neighboring directional modes
1100 if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
1101 if (conditional_skipintra(mode, *best_mode)) continue;
1104 memcpy(tempa, ta, num_4x4_blocks_wide * sizeof(ta[0]));
1105 memcpy(templ, tl, num_4x4_blocks_high * sizeof(tl[0]));
1107 for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
1108 for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
1109 const int block = (row + idy) * 2 + (col + idx);
1110 const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
1111 uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
1112 int16_t *const src_diff =
1113 vp9_raster_block_offset_int16(BLOCK_8X8, block, p->src_diff);
1114 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
1115 xd->mi[0]->bmi[block].as_mode = mode;
1116 vp9_predict_intra_block(xd, 1, TX_4X4, mode, x->skip_encode ? src : dst,
1117 x->skip_encode ? src_stride : dst_stride, dst,
1118 dst_stride, col + idx, row + idy, 0);
1119 vpx_subtract_block(4, 4, src_diff, 8, src, src_stride, dst, dst_stride);
1122 const scan_order *so = &vp9_default_scan_orders[TX_4X4];
1123 const int coeff_ctx =
1124 combine_entropy_contexts(tempa[idx], templ[idy]);
1125 vp9_fwht4x4(src_diff, coeff, 8);
1126 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
1127 ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
1128 so->neighbors, cpi->sf.use_fast_coef_costing);
1129 tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0) ? 1 : 0;
1130 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
1132 vp9_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block), dst, dst_stride,
1136 const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
1137 const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
1138 const int coeff_ctx =
1139 combine_entropy_contexts(tempa[idx], templ[idy]);
1140 vp9_fht4x4(src_diff, coeff, 8, tx_type);
1141 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
1142 ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
1143 so->neighbors, cpi->sf.use_fast_coef_costing);
1144 tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0) ? 1 : 0;
1145 distortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, block),
1148 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
1150 vp9_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block), dst,
1151 dst_stride, p->eobs[block]);
1157 this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
1159 if (this_rd < best_rd) {
1162 *bestdistortion = distortion;
1165 memcpy(a, tempa, num_4x4_blocks_wide * sizeof(tempa[0]));
1166 memcpy(l, templ, num_4x4_blocks_high * sizeof(templ[0]));
1167 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
1168 memcpy(best_dst + idy * 8, dst_init + idy * dst_stride,
1169 num_4x4_blocks_wide * 4);
1174 if (best_rd >= rd_thresh || x->skip_encode) return best_rd;
1176 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
1177 memcpy(dst_init + idy * dst_stride, best_dst + idy * 8,
1178 num_4x4_blocks_wide * 4);
1183 static int64_t rd_pick_intra_sub_8x8_y_mode(VP9_COMP *cpi, MACROBLOCK *mb,
1184 int *rate, int *rate_y,
1185 int64_t *distortion,
1188 const MACROBLOCKD *const xd = &mb->e_mbd;
1189 MODE_INFO *const mic = xd->mi[0];
1190 const MODE_INFO *above_mi = xd->above_mi;
1191 const MODE_INFO *left_mi = xd->left_mi;
1192 const BLOCK_SIZE bsize = xd->mi[0]->sb_type;
1193 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1194 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1197 int64_t total_distortion = 0;
1199 int64_t total_rd = 0;
1200 const int *bmode_costs = cpi->mbmode_cost;
1202 // Pick modes for each sub-block (of size 4x4, 4x8, or 8x4) in an 8x8 block.
1203 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1204 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1205 PREDICTION_MODE best_mode = DC_PRED;
1206 int r = INT_MAX, ry = INT_MAX;
1207 int64_t d = INT64_MAX, this_rd = INT64_MAX;
1209 if (cpi->common.frame_type == KEY_FRAME) {
1210 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, i);
1211 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, i);
1213 bmode_costs = cpi->y_mode_costs[A][L];
1216 this_rd = rd_pick_intra4x4block(
1217 cpi, mb, idy, idx, &best_mode, bmode_costs,
1218 xd->plane[0].above_context + idx, xd->plane[0].left_context + idy, &r,
1219 &ry, &d, bsize, best_rd - total_rd);
1221 if (this_rd >= best_rd - total_rd) return INT64_MAX;
1223 total_rd += this_rd;
1225 total_distortion += d;
1228 mic->bmi[i].as_mode = best_mode;
1229 for (j = 1; j < num_4x4_blocks_high; ++j)
1230 mic->bmi[i + j * 2].as_mode = best_mode;
1231 for (j = 1; j < num_4x4_blocks_wide; ++j)
1232 mic->bmi[i + j].as_mode = best_mode;
1234 if (total_rd >= best_rd) return INT64_MAX;
1239 *rate_y = tot_rate_y;
1240 *distortion = total_distortion;
1241 mic->mode = mic->bmi[3].as_mode;
1243 return RDCOST(mb->rdmult, mb->rddiv, cost, total_distortion);
1246 // This function is used only for intra_only frames
1247 static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
1248 int *rate_tokenonly, int64_t *distortion,
1249 int *skippable, BLOCK_SIZE bsize,
1251 PREDICTION_MODE mode;
1252 PREDICTION_MODE mode_selected = DC_PRED;
1253 MACROBLOCKD *const xd = &x->e_mbd;
1254 MODE_INFO *const mic = xd->mi[0];
1255 int this_rate, this_rate_tokenonly, s;
1256 int64_t this_distortion, this_rd;
1257 TX_SIZE best_tx = TX_4X4;
1259 const MODE_INFO *above_mi = xd->above_mi;
1260 const MODE_INFO *left_mi = xd->left_mi;
1261 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
1262 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
1263 bmode_costs = cpi->y_mode_costs[A][L];
1265 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1266 /* Y Search for intra prediction mode */
1267 for (mode = DC_PRED; mode <= TM_PRED; mode++) {
1268 if (cpi->sf.use_nonrd_pick_mode) {
1269 // These speed features are turned on in hybrid non-RD and RD mode
1270 // for key frame coding in the context of real-time setting.
1271 if (conditional_skipintra(mode, mode_selected)) continue;
1272 if (*skippable) break;
1277 super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion, &s, NULL,
1280 if (this_rate_tokenonly == INT_MAX) continue;
1282 this_rate = this_rate_tokenonly + bmode_costs[mode];
1283 this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1285 if (this_rd < best_rd) {
1286 mode_selected = mode;
1288 best_tx = mic->tx_size;
1290 *rate_tokenonly = this_rate_tokenonly;
1291 *distortion = this_distortion;
1296 mic->mode = mode_selected;
1297 mic->tx_size = best_tx;
1302 // Return value 0: early termination triggered, no valid rd cost available;
1303 // 1: rd cost values are valid.
1304 static int super_block_uvrd(const VP9_COMP *cpi, MACROBLOCK *x, int *rate,
1305 int64_t *distortion, int *skippable, int64_t *sse,
1306 BLOCK_SIZE bsize, int64_t ref_best_rd) {
1307 MACROBLOCKD *const xd = &x->e_mbd;
1308 MODE_INFO *const mi = xd->mi[0];
1309 const TX_SIZE uv_tx_size = get_uv_tx_size(mi, &xd->plane[1]);
1311 int pnrate = 0, pnskip = 1;
1312 int64_t pndist = 0, pnsse = 0;
1313 int is_cost_valid = 1;
1315 if (ref_best_rd < 0) is_cost_valid = 0;
1317 if (is_inter_block(mi) && is_cost_valid) {
1319 for (plane = 1; plane < MAX_MB_PLANE; ++plane)
1320 vp9_subtract_plane(x, bsize, plane);
1328 for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
1329 txfm_rd_in_plane(cpi, x, &pnrate, &pndist, &pnskip, &pnsse, ref_best_rd,
1330 plane, bsize, uv_tx_size, cpi->sf.use_fast_coef_costing);
1331 if (pnrate == INT_MAX) {
1336 *distortion += pndist;
1338 *skippable &= pnskip;
1341 if (!is_cost_valid) {
1344 *distortion = INT64_MAX;
1349 return is_cost_valid;
1352 static int64_t rd_pick_intra_sbuv_mode(VP9_COMP *cpi, MACROBLOCK *x,
1353 PICK_MODE_CONTEXT *ctx, int *rate,
1354 int *rate_tokenonly, int64_t *distortion,
1355 int *skippable, BLOCK_SIZE bsize,
1356 TX_SIZE max_tx_size) {
1357 MACROBLOCKD *xd = &x->e_mbd;
1358 PREDICTION_MODE mode;
1359 PREDICTION_MODE mode_selected = DC_PRED;
1360 int64_t best_rd = INT64_MAX, this_rd;
1361 int this_rate_tokenonly, this_rate, s;
1362 int64_t this_distortion, this_sse;
1364 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1365 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
1366 if (!(cpi->sf.intra_uv_mode_mask[max_tx_size] & (1 << mode))) continue;
1367 #if CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
1368 if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) &&
1369 (xd->above_mi == NULL || xd->left_mi == NULL) && need_top_left[mode])
1371 #endif // CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
1373 xd->mi[0]->uv_mode = mode;
1375 if (!super_block_uvrd(cpi, x, &this_rate_tokenonly, &this_distortion, &s,
1376 &this_sse, bsize, best_rd))
1379 this_rate_tokenonly +
1380 cpi->intra_uv_mode_cost[cpi->common.frame_type][xd->mi[0]->mode][mode];
1381 this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1383 if (this_rd < best_rd) {
1384 mode_selected = mode;
1387 *rate_tokenonly = this_rate_tokenonly;
1388 *distortion = this_distortion;
1390 if (!x->select_tx_size) swap_block_ptr(x, ctx, 2, 0, 1, MAX_MB_PLANE);
1394 xd->mi[0]->uv_mode = mode_selected;
1398 static int64_t rd_sbuv_dcpred(const VP9_COMP *cpi, MACROBLOCK *x, int *rate,
1399 int *rate_tokenonly, int64_t *distortion,
1400 int *skippable, BLOCK_SIZE bsize) {
1401 const VP9_COMMON *cm = &cpi->common;
1404 x->e_mbd.mi[0]->uv_mode = DC_PRED;
1405 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1406 super_block_uvrd(cpi, x, rate_tokenonly, distortion, skippable, &unused,
1410 cpi->intra_uv_mode_cost[cm->frame_type][x->e_mbd.mi[0]->mode][DC_PRED];
1411 return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
1414 static void choose_intra_uv_mode(VP9_COMP *cpi, MACROBLOCK *const x,
1415 PICK_MODE_CONTEXT *ctx, BLOCK_SIZE bsize,
1416 TX_SIZE max_tx_size, int *rate_uv,
1417 int *rate_uv_tokenonly, int64_t *dist_uv,
1418 int *skip_uv, PREDICTION_MODE *mode_uv) {
1419 // Use an estimated rd for uv_intra based on DC_PRED if the
1420 // appropriate speed flag is set.
1421 if (cpi->sf.use_uv_intra_rd_estimate) {
1422 rd_sbuv_dcpred(cpi, x, rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
1423 bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize);
1424 // Else do a proper rd search for each possible transform size that may
1425 // be considered in the main rd loop.
1427 rd_pick_intra_sbuv_mode(cpi, x, ctx, rate_uv, rate_uv_tokenonly, dist_uv,
1428 skip_uv, bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize,
1431 *mode_uv = x->e_mbd.mi[0]->uv_mode;
1434 static int cost_mv_ref(const VP9_COMP *cpi, PREDICTION_MODE mode,
1436 assert(is_inter_mode(mode));
1437 return cpi->inter_mode_cost[mode_context][INTER_OFFSET(mode)];
1440 static int set_and_cost_bmi_mvs(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
1441 int i, PREDICTION_MODE mode, int_mv this_mv[2],
1442 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1443 int_mv seg_mvs[MAX_REF_FRAMES],
1444 int_mv *best_ref_mv[2], const int *mvjcost,
1446 MODE_INFO *const mi = xd->mi[0];
1447 const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1450 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[mi->sb_type];
1451 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mi->sb_type];
1452 const int is_compound = has_second_ref(mi);
1456 this_mv[0].as_int = seg_mvs[mi->ref_frame[0]].as_int;
1457 thismvcost += vp9_mv_bit_cost(&this_mv[0].as_mv, &best_ref_mv[0]->as_mv,
1458 mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1460 this_mv[1].as_int = seg_mvs[mi->ref_frame[1]].as_int;
1461 thismvcost += vp9_mv_bit_cost(&this_mv[1].as_mv, &best_ref_mv[1]->as_mv,
1462 mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1467 this_mv[0].as_int = frame_mv[mode][mi->ref_frame[0]].as_int;
1469 this_mv[1].as_int = frame_mv[mode][mi->ref_frame[1]].as_int;
1472 this_mv[0].as_int = 0;
1473 if (is_compound) this_mv[1].as_int = 0;
1478 mi->bmi[i].as_mv[0].as_int = this_mv[0].as_int;
1479 if (is_compound) mi->bmi[i].as_mv[1].as_int = this_mv[1].as_int;
1481 mi->bmi[i].as_mode = mode;
1483 for (idy = 0; idy < num_4x4_blocks_high; ++idy)
1484 for (idx = 0; idx < num_4x4_blocks_wide; ++idx)
1485 memmove(&mi->bmi[i + idy * 2 + idx], &mi->bmi[i], sizeof(mi->bmi[i]));
1487 return cost_mv_ref(cpi, mode, mbmi_ext->mode_context[mi->ref_frame[0]]) +
1491 static int64_t encode_inter_mb_segment(VP9_COMP *cpi, MACROBLOCK *x,
1492 int64_t best_yrd, int i, int *labelyrate,
1493 int64_t *distortion, int64_t *sse,
1494 ENTROPY_CONTEXT *ta, ENTROPY_CONTEXT *tl,
1495 int mi_row, int mi_col) {
1497 MACROBLOCKD *xd = &x->e_mbd;
1498 struct macroblockd_plane *const pd = &xd->plane[0];
1499 struct macroblock_plane *const p = &x->plane[0];
1500 MODE_INFO *const mi = xd->mi[0];
1501 const BLOCK_SIZE plane_bsize = get_plane_block_size(mi->sb_type, pd);
1502 const int width = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
1503 const int height = 4 * num_4x4_blocks_high_lookup[plane_bsize];
1506 const uint8_t *const src =
1507 &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
1508 uint8_t *const dst =
1509 &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)];
1510 int64_t thisdistortion = 0, thissse = 0;
1511 int thisrate = 0, ref;
1512 const scan_order *so = &vp9_default_scan_orders[TX_4X4];
1513 const int is_compound = has_second_ref(mi);
1514 const InterpKernel *kernel = vp9_filter_kernels[mi->interp_filter];
1516 for (ref = 0; ref < 1 + is_compound; ++ref) {
1517 const int bw = b_width_log2_lookup[BLOCK_8X8];
1518 const int h = 4 * (i >> bw);
1519 const int w = 4 * (i & ((1 << bw) - 1));
1520 const struct scale_factors *sf = &xd->block_refs[ref]->sf;
1521 int y_stride = pd->pre[ref].stride;
1522 uint8_t *pre = pd->pre[ref].buf + (h * pd->pre[ref].stride + w);
1524 if (vp9_is_scaled(sf)) {
1525 const int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
1526 const int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
1528 y_stride = xd->block_refs[ref]->buf->y_stride;
1529 pre = xd->block_refs[ref]->buf->y_buffer;
1530 pre += scaled_buffer_offset(x_start + w, y_start + h, y_stride, sf);
1532 #if CONFIG_VP9_HIGHBITDEPTH
1533 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1534 vp9_highbd_build_inter_predictor(
1535 CONVERT_TO_SHORTPTR(pre), y_stride, CONVERT_TO_SHORTPTR(dst),
1536 pd->dst.stride, &mi->bmi[i].as_mv[ref].as_mv,
1537 &xd->block_refs[ref]->sf, width, height, ref, kernel, MV_PRECISION_Q3,
1538 mi_col * MI_SIZE + 4 * (i % 2), mi_row * MI_SIZE + 4 * (i / 2),
1541 vp9_build_inter_predictor(
1542 pre, y_stride, dst, pd->dst.stride, &mi->bmi[i].as_mv[ref].as_mv,
1543 &xd->block_refs[ref]->sf, width, height, ref, kernel, MV_PRECISION_Q3,
1544 mi_col * MI_SIZE + 4 * (i % 2), mi_row * MI_SIZE + 4 * (i / 2));
1547 vp9_build_inter_predictor(
1548 pre, y_stride, dst, pd->dst.stride, &mi->bmi[i].as_mv[ref].as_mv,
1549 &xd->block_refs[ref]->sf, width, height, ref, kernel, MV_PRECISION_Q3,
1550 mi_col * MI_SIZE + 4 * (i % 2), mi_row * MI_SIZE + 4 * (i / 2));
1551 #endif // CONFIG_VP9_HIGHBITDEPTH
1554 #if CONFIG_VP9_HIGHBITDEPTH
1555 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1556 vpx_highbd_subtract_block(
1557 height, width, vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1558 8, src, p->src.stride, dst, pd->dst.stride, xd->bd);
1560 vpx_subtract_block(height, width,
1561 vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1562 8, src, p->src.stride, dst, pd->dst.stride);
1565 vpx_subtract_block(height, width,
1566 vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1567 8, src, p->src.stride, dst, pd->dst.stride);
1568 #endif // CONFIG_VP9_HIGHBITDEPTH
1571 for (idy = 0; idy < height / 4; ++idy) {
1572 for (idx = 0; idx < width / 4; ++idx) {
1573 #if CONFIG_VP9_HIGHBITDEPTH
1574 const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
1576 int64_t ssz, rd, rd1, rd2;
1579 k += (idy * 2 + idx);
1580 coeff_ctx = combine_entropy_contexts(ta[k & 1], tl[k >> 1]);
1581 coeff = BLOCK_OFFSET(p->coeff, k);
1582 x->fwd_txfm4x4(vp9_raster_block_offset_int16(BLOCK_8X8, k, p->src_diff),
1584 vp9_regular_quantize_b_4x4(x, 0, k, so->scan, so->iscan);
1585 #if CONFIG_VP9_HIGHBITDEPTH
1586 thisdistortion += vp9_highbd_block_error_dispatch(
1587 coeff, BLOCK_OFFSET(pd->dqcoeff, k), 16, &ssz, bd);
1590 vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k), 16, &ssz);
1591 #endif // CONFIG_VP9_HIGHBITDEPTH
1593 thisrate += cost_coeffs(x, 0, k, TX_4X4, coeff_ctx, so->scan,
1594 so->neighbors, cpi->sf.use_fast_coef_costing);
1595 ta[k & 1] = tl[k >> 1] = (x->plane[0].eobs[k] > 0) ? 1 : 0;
1596 rd1 = RDCOST(x->rdmult, x->rddiv, thisrate, thisdistortion >> 2);
1597 rd2 = RDCOST(x->rdmult, x->rddiv, 0, thissse >> 2);
1598 rd = VPXMIN(rd1, rd2);
1599 if (rd >= best_yrd) return INT64_MAX;
1603 *distortion = thisdistortion >> 2;
1604 *labelyrate = thisrate;
1605 *sse = thissse >> 2;
1607 return RDCOST(x->rdmult, x->rddiv, *labelyrate, *distortion);
1618 ENTROPY_CONTEXT ta[2];
1619 ENTROPY_CONTEXT tl[2];
1631 PREDICTION_MODE modes[4];
1632 SEG_RDSTAT rdstat[4][INTER_MODES];
1636 static INLINE int mv_check_bounds(const MvLimits *mv_limits, const MV *mv) {
1637 return (mv->row >> 3) < mv_limits->row_min ||
1638 (mv->row >> 3) > mv_limits->row_max ||
1639 (mv->col >> 3) < mv_limits->col_min ||
1640 (mv->col >> 3) > mv_limits->col_max;
1643 static INLINE void mi_buf_shift(MACROBLOCK *x, int i) {
1644 MODE_INFO *const mi = x->e_mbd.mi[0];
1645 struct macroblock_plane *const p = &x->plane[0];
1646 struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
1649 &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
1650 assert(((intptr_t)pd->pre[0].buf & 0x7) == 0);
1652 &pd->pre[0].buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->pre[0].stride)];
1653 if (has_second_ref(mi))
1656 .buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->pre[1].stride)];
1659 static INLINE void mi_buf_restore(MACROBLOCK *x, struct buf_2d orig_src,
1660 struct buf_2d orig_pre[2]) {
1661 MODE_INFO *mi = x->e_mbd.mi[0];
1662 x->plane[0].src = orig_src;
1663 x->e_mbd.plane[0].pre[0] = orig_pre[0];
1664 if (has_second_ref(mi)) x->e_mbd.plane[0].pre[1] = orig_pre[1];
1667 static INLINE int mv_has_subpel(const MV *mv) {
1668 return (mv->row & 0x0F) || (mv->col & 0x0F);
1671 // Check if NEARESTMV/NEARMV/ZEROMV is the cheapest way encode zero motion.
1672 // TODO(aconverse): Find out if this is still productive then clean up or remove
1673 static int check_best_zero_mv(const VP9_COMP *cpi,
1674 const uint8_t mode_context[MAX_REF_FRAMES],
1675 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1677 const MV_REFERENCE_FRAME ref_frames[2]) {
1678 if ((this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) &&
1679 frame_mv[this_mode][ref_frames[0]].as_int == 0 &&
1680 (ref_frames[1] == NONE ||
1681 frame_mv[this_mode][ref_frames[1]].as_int == 0)) {
1682 int rfc = mode_context[ref_frames[0]];
1683 int c1 = cost_mv_ref(cpi, NEARMV, rfc);
1684 int c2 = cost_mv_ref(cpi, NEARESTMV, rfc);
1685 int c3 = cost_mv_ref(cpi, ZEROMV, rfc);
1687 if (this_mode == NEARMV) {
1688 if (c1 > c3) return 0;
1689 } else if (this_mode == NEARESTMV) {
1690 if (c2 > c3) return 0;
1692 assert(this_mode == ZEROMV);
1693 if (ref_frames[1] == NONE) {
1694 if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0) ||
1695 (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0))
1698 if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0 &&
1699 frame_mv[NEARESTMV][ref_frames[1]].as_int == 0) ||
1700 (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0 &&
1701 frame_mv[NEARMV][ref_frames[1]].as_int == 0))
1709 static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
1710 int_mv *frame_mv, int mi_row, int mi_col,
1711 int_mv single_newmv[MAX_REF_FRAMES],
1713 const VP9_COMMON *const cm = &cpi->common;
1714 const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
1715 const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
1716 MACROBLOCKD *xd = &x->e_mbd;
1717 MODE_INFO *mi = xd->mi[0];
1718 const int refs[2] = { mi->ref_frame[0],
1719 mi->ref_frame[1] < 0 ? 0 : mi->ref_frame[1] };
1722 const InterpKernel *kernel = vp9_filter_kernels[mi->interp_filter];
1723 struct scale_factors sf;
1725 // Do joint motion search in compound mode to get more accurate mv.
1726 struct buf_2d backup_yv12[2][MAX_MB_PLANE];
1727 uint32_t last_besterr[2] = { UINT_MAX, UINT_MAX };
1728 const YV12_BUFFER_CONFIG *const scaled_ref_frame[2] = {
1729 vp9_get_scaled_ref_frame(cpi, mi->ref_frame[0]),
1730 vp9_get_scaled_ref_frame(cpi, mi->ref_frame[1])
1733 // Prediction buffer from second frame.
1734 #if CONFIG_VP9_HIGHBITDEPTH
1735 DECLARE_ALIGNED(16, uint16_t, second_pred_alloc_16[64 * 64]);
1736 uint8_t *second_pred;
1738 DECLARE_ALIGNED(16, uint8_t, second_pred[64 * 64]);
1739 #endif // CONFIG_VP9_HIGHBITDEPTH
1741 for (ref = 0; ref < 2; ++ref) {
1742 ref_mv[ref] = x->mbmi_ext->ref_mvs[refs[ref]][0];
1744 if (scaled_ref_frame[ref]) {
1746 // Swap out the reference frame for a version that's been scaled to
1747 // match the resolution of the current frame, allowing the existing
1748 // motion search code to be used without additional modifications.
1749 for (i = 0; i < MAX_MB_PLANE; i++)
1750 backup_yv12[ref][i] = xd->plane[i].pre[ref];
1751 vp9_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
1755 frame_mv[refs[ref]].as_int = single_newmv[refs[ref]].as_int;
1758 // Since we have scaled the reference frames to match the size of the current
1759 // frame we must use a unit scaling factor during mode selection.
1760 #if CONFIG_VP9_HIGHBITDEPTH
1761 vp9_setup_scale_factors_for_frame(&sf, cm->width, cm->height, cm->width,
1762 cm->height, cm->use_highbitdepth);
1764 vp9_setup_scale_factors_for_frame(&sf, cm->width, cm->height, cm->width,
1766 #endif // CONFIG_VP9_HIGHBITDEPTH
1768 // Allow joint search multiple times iteratively for each reference frame
1769 // and break out of the search loop if it couldn't find a better mv.
1770 for (ite = 0; ite < 4; ite++) {
1771 struct buf_2d ref_yv12[2];
1772 uint32_t bestsme = UINT_MAX;
1773 int sadpb = x->sadperbit16;
1775 int search_range = 3;
1777 const MvLimits tmp_mv_limits = x->mv_limits;
1778 int id = ite % 2; // Even iterations search in the first reference frame,
1779 // odd iterations search in the second. The predictor
1780 // found for the 'other' reference frame is factored in.
1782 // Initialized here because of compiler problem in Visual Studio.
1783 ref_yv12[0] = xd->plane[0].pre[0];
1784 ref_yv12[1] = xd->plane[0].pre[1];
1786 // Get the prediction block from the 'other' reference frame.
1787 #if CONFIG_VP9_HIGHBITDEPTH
1788 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1789 second_pred = CONVERT_TO_BYTEPTR(second_pred_alloc_16);
1790 vp9_highbd_build_inter_predictor(
1791 CONVERT_TO_SHORTPTR(ref_yv12[!id].buf), ref_yv12[!id].stride,
1792 second_pred_alloc_16, pw, &frame_mv[refs[!id]].as_mv, &sf, pw, ph, 0,
1793 kernel, MV_PRECISION_Q3, mi_col * MI_SIZE, mi_row * MI_SIZE, xd->bd);
1795 second_pred = (uint8_t *)second_pred_alloc_16;
1796 vp9_build_inter_predictor(ref_yv12[!id].buf, ref_yv12[!id].stride,
1797 second_pred, pw, &frame_mv[refs[!id]].as_mv,
1798 &sf, pw, ph, 0, kernel, MV_PRECISION_Q3,
1799 mi_col * MI_SIZE, mi_row * MI_SIZE);
1802 vp9_build_inter_predictor(ref_yv12[!id].buf, ref_yv12[!id].stride,
1803 second_pred, pw, &frame_mv[refs[!id]].as_mv, &sf,
1804 pw, ph, 0, kernel, MV_PRECISION_Q3,
1805 mi_col * MI_SIZE, mi_row * MI_SIZE);
1806 #endif // CONFIG_VP9_HIGHBITDEPTH
1808 // Do compound motion search on the current reference frame.
1809 if (id) xd->plane[0].pre[0] = ref_yv12[id];
1810 vp9_set_mv_search_range(&x->mv_limits, &ref_mv[id].as_mv);
1812 // Use the mv result from the single mode as mv predictor.
1813 tmp_mv = frame_mv[refs[id]].as_mv;
1818 // Small-range full-pixel motion search.
1819 bestsme = vp9_refining_search_8p_c(x, &tmp_mv, sadpb, search_range,
1820 &cpi->fn_ptr[bsize], &ref_mv[id].as_mv,
1822 if (bestsme < UINT_MAX)
1823 bestsme = vp9_get_mvpred_av_var(x, &tmp_mv, &ref_mv[id].as_mv,
1824 second_pred, &cpi->fn_ptr[bsize], 1);
1826 x->mv_limits = tmp_mv_limits;
1828 if (bestsme < UINT_MAX) {
1829 uint32_t dis; /* TODO: use dis in distortion calculation later. */
1831 bestsme = cpi->find_fractional_mv_step(
1832 x, &tmp_mv, &ref_mv[id].as_mv, cpi->common.allow_high_precision_mv,
1833 x->errorperbit, &cpi->fn_ptr[bsize], 0,
1834 cpi->sf.mv.subpel_iters_per_step, NULL, x->nmvjointcost, x->mvcost,
1835 &dis, &sse, second_pred, pw, ph);
1838 // Restore the pointer to the first (possibly scaled) prediction buffer.
1839 if (id) xd->plane[0].pre[0] = ref_yv12[0];
1841 if (bestsme < last_besterr[id]) {
1842 frame_mv[refs[id]].as_mv = tmp_mv;
1843 last_besterr[id] = bestsme;
1851 for (ref = 0; ref < 2; ++ref) {
1852 if (scaled_ref_frame[ref]) {
1853 // Restore the prediction frame pointers to their unscaled versions.
1855 for (i = 0; i < MAX_MB_PLANE; i++)
1856 xd->plane[i].pre[ref] = backup_yv12[ref][i];
1859 *rate_mv += vp9_mv_bit_cost(&frame_mv[refs[ref]].as_mv,
1860 &x->mbmi_ext->ref_mvs[refs[ref]][0].as_mv,
1861 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1865 static int64_t rd_pick_best_sub8x8_mode(
1866 VP9_COMP *cpi, MACROBLOCK *x, int_mv *best_ref_mv,
1867 int_mv *second_best_ref_mv, int64_t best_rd, int *returntotrate,
1868 int *returnyrate, int64_t *returndistortion, int *skippable, int64_t *psse,
1869 int mvthresh, int_mv seg_mvs[4][MAX_REF_FRAMES], BEST_SEG_INFO *bsi_buf,
1870 int filter_idx, int mi_row, int mi_col) {
1872 BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
1873 MACROBLOCKD *xd = &x->e_mbd;
1874 MODE_INFO *mi = xd->mi[0];
1876 int k, br = 0, idx, idy;
1877 int64_t bd = 0, block_sse = 0;
1878 PREDICTION_MODE this_mode;
1879 VP9_COMMON *cm = &cpi->common;
1880 struct macroblock_plane *const p = &x->plane[0];
1881 struct macroblockd_plane *const pd = &xd->plane[0];
1882 const int label_count = 4;
1883 int64_t this_segment_rd = 0;
1884 int label_mv_thresh;
1885 int segmentyrate = 0;
1886 const BLOCK_SIZE bsize = mi->sb_type;
1887 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1888 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1889 ENTROPY_CONTEXT t_above[2], t_left[2];
1890 int subpelmv = 1, have_ref = 0;
1891 SPEED_FEATURES *const sf = &cpi->sf;
1892 const int has_second_rf = has_second_ref(mi);
1893 const int inter_mode_mask = sf->inter_mode_mask[bsize];
1894 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1898 bsi->segment_rd = best_rd;
1899 bsi->ref_mv[0] = best_ref_mv;
1900 bsi->ref_mv[1] = second_best_ref_mv;
1901 bsi->mvp.as_int = best_ref_mv->as_int;
1902 bsi->mvthresh = mvthresh;
1904 for (i = 0; i < 4; i++) bsi->modes[i] = ZEROMV;
1906 memcpy(t_above, pd->above_context, sizeof(t_above));
1907 memcpy(t_left, pd->left_context, sizeof(t_left));
1909 // 64 makes this threshold really big effectively
1910 // making it so that we very rarely check mvs on
1911 // segments. setting this to 1 would make mv thresh
1912 // roughly equal to what it is for macroblocks
1913 label_mv_thresh = 1 * bsi->mvthresh / label_count;
1915 // Segmentation method overheads
1916 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1917 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1918 // TODO(jingning,rbultje): rewrite the rate-distortion optimization
1919 // loop for 4x4/4x8/8x4 block coding. to be replaced with new rd loop
1920 int_mv mode_mv[MB_MODE_COUNT][2];
1921 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1922 PREDICTION_MODE mode_selected = ZEROMV;
1923 int64_t best_rd = INT64_MAX;
1924 const int i = idy * 2 + idx;
1927 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1928 const MV_REFERENCE_FRAME frame = mi->ref_frame[ref];
1929 frame_mv[ZEROMV][frame].as_int = 0;
1930 vp9_append_sub8x8_mvs_for_idx(
1931 cm, xd, i, ref, mi_row, mi_col, &frame_mv[NEARESTMV][frame],
1932 &frame_mv[NEARMV][frame], mbmi_ext->mode_context);
1935 // search for the best motion vector on this segment
1936 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
1937 const struct buf_2d orig_src = x->plane[0].src;
1938 struct buf_2d orig_pre[2];
1940 mode_idx = INTER_OFFSET(this_mode);
1941 bsi->rdstat[i][mode_idx].brdcost = INT64_MAX;
1942 if (!(inter_mode_mask & (1 << this_mode))) continue;
1944 if (!check_best_zero_mv(cpi, mbmi_ext->mode_context, frame_mv,
1945 this_mode, mi->ref_frame))
1948 memcpy(orig_pre, pd->pre, sizeof(orig_pre));
1949 memcpy(bsi->rdstat[i][mode_idx].ta, t_above,
1950 sizeof(bsi->rdstat[i][mode_idx].ta));
1951 memcpy(bsi->rdstat[i][mode_idx].tl, t_left,
1952 sizeof(bsi->rdstat[i][mode_idx].tl));
1954 // motion search for newmv (single predictor case only)
1955 if (!has_second_rf && this_mode == NEWMV &&
1956 seg_mvs[i][mi->ref_frame[0]].as_int == INVALID_MV) {
1957 MV *const new_mv = &mode_mv[NEWMV][0].as_mv;
1959 uint32_t bestsme = UINT_MAX;
1960 int sadpb = x->sadperbit4;
1964 const MvLimits tmp_mv_limits = x->mv_limits;
1966 /* Is the best so far sufficiently good that we cant justify doing
1967 * and new motion search. */
1968 if (best_rd < label_mv_thresh) break;
1970 if (cpi->oxcf.mode != BEST) {
1971 // use previous block's result as next block's MV predictor.
1973 bsi->mvp.as_int = mi->bmi[i - 1].as_mv[0].as_int;
1974 if (i == 2) bsi->mvp.as_int = mi->bmi[i - 2].as_mv[0].as_int;
1978 max_mv = x->max_mv_context[mi->ref_frame[0]];
1981 VPXMAX(abs(bsi->mvp.as_mv.row), abs(bsi->mvp.as_mv.col)) >> 3;
1983 if (sf->mv.auto_mv_step_size && cm->show_frame) {
1984 // Take wtd average of the step_params based on the last frame's
1985 // max mv magnitude and the best ref mvs of the current block for
1986 // the given reference.
1988 (vp9_init_search_range(max_mv) + cpi->mv_step_param) / 2;
1990 step_param = cpi->mv_step_param;
1993 mvp_full.row = bsi->mvp.as_mv.row >> 3;
1994 mvp_full.col = bsi->mvp.as_mv.col >> 3;
1996 if (sf->adaptive_motion_search) {
1997 mvp_full.row = x->pred_mv[mi->ref_frame[0]].row >> 3;
1998 mvp_full.col = x->pred_mv[mi->ref_frame[0]].col >> 3;
1999 step_param = VPXMAX(step_param, 8);
2002 // adjust src pointer for this block
2005 vp9_set_mv_search_range(&x->mv_limits, &bsi->ref_mv[0]->as_mv);
2007 bestsme = vp9_full_pixel_search(
2008 cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method,
2010 sf->mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL,
2011 &bsi->ref_mv[0]->as_mv, new_mv, INT_MAX, 1);
2013 x->mv_limits = tmp_mv_limits;
2015 if (bestsme < UINT_MAX) {
2016 uint32_t distortion;
2017 cpi->find_fractional_mv_step(
2018 x, new_mv, &bsi->ref_mv[0]->as_mv, cm->allow_high_precision_mv,
2019 x->errorperbit, &cpi->fn_ptr[bsize], sf->mv.subpel_force_stop,
2020 sf->mv.subpel_iters_per_step, cond_cost_list(cpi, cost_list),
2021 x->nmvjointcost, x->mvcost, &distortion,
2022 &x->pred_sse[mi->ref_frame[0]], NULL, 0, 0);
2024 // save motion search result for use in compound prediction
2025 seg_mvs[i][mi->ref_frame[0]].as_mv = *new_mv;
2028 if (sf->adaptive_motion_search)
2029 x->pred_mv[mi->ref_frame[0]] = *new_mv;
2031 // restore src pointers
2032 mi_buf_restore(x, orig_src, orig_pre);
2035 if (has_second_rf) {
2036 if (seg_mvs[i][mi->ref_frame[1]].as_int == INVALID_MV ||
2037 seg_mvs[i][mi->ref_frame[0]].as_int == INVALID_MV)
2041 if (has_second_rf && this_mode == NEWMV &&
2042 mi->interp_filter == EIGHTTAP) {
2043 // adjust src pointers
2045 if (sf->comp_inter_joint_search_thresh <= bsize) {
2047 joint_motion_search(cpi, x, bsize, frame_mv[this_mode], mi_row,
2048 mi_col, seg_mvs[i], &rate_mv);
2049 seg_mvs[i][mi->ref_frame[0]].as_int =
2050 frame_mv[this_mode][mi->ref_frame[0]].as_int;
2051 seg_mvs[i][mi->ref_frame[1]].as_int =
2052 frame_mv[this_mode][mi->ref_frame[1]].as_int;
2054 // restore src pointers
2055 mi_buf_restore(x, orig_src, orig_pre);
2058 bsi->rdstat[i][mode_idx].brate = set_and_cost_bmi_mvs(
2059 cpi, x, xd, i, this_mode, mode_mv[this_mode], frame_mv, seg_mvs[i],
2060 bsi->ref_mv, x->nmvjointcost, x->mvcost);
2062 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
2063 bsi->rdstat[i][mode_idx].mvs[ref].as_int =
2064 mode_mv[this_mode][ref].as_int;
2065 if (num_4x4_blocks_wide > 1)
2066 bsi->rdstat[i + 1][mode_idx].mvs[ref].as_int =
2067 mode_mv[this_mode][ref].as_int;
2068 if (num_4x4_blocks_high > 1)
2069 bsi->rdstat[i + 2][mode_idx].mvs[ref].as_int =
2070 mode_mv[this_mode][ref].as_int;
2073 // Trap vectors that reach beyond the UMV borders
2074 if (mv_check_bounds(&x->mv_limits, &mode_mv[this_mode][0].as_mv) ||
2076 mv_check_bounds(&x->mv_limits, &mode_mv[this_mode][1].as_mv)))
2079 if (filter_idx > 0) {
2080 BEST_SEG_INFO *ref_bsi = bsi_buf;
2084 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
2085 subpelmv |= mv_has_subpel(&mode_mv[this_mode][ref].as_mv);
2086 have_ref &= mode_mv[this_mode][ref].as_int ==
2087 ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
2090 if (filter_idx > 1 && !subpelmv && !have_ref) {
2091 ref_bsi = bsi_buf + 1;
2093 for (ref = 0; ref < 1 + has_second_rf; ++ref)
2094 have_ref &= mode_mv[this_mode][ref].as_int ==
2095 ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
2098 if (!subpelmv && have_ref &&
2099 ref_bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
2100 memcpy(&bsi->rdstat[i][mode_idx], &ref_bsi->rdstat[i][mode_idx],
2101 sizeof(SEG_RDSTAT));
2102 if (num_4x4_blocks_wide > 1)
2103 bsi->rdstat[i + 1][mode_idx].eobs =
2104 ref_bsi->rdstat[i + 1][mode_idx].eobs;
2105 if (num_4x4_blocks_high > 1)
2106 bsi->rdstat[i + 2][mode_idx].eobs =
2107 ref_bsi->rdstat[i + 2][mode_idx].eobs;
2109 if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
2110 mode_selected = this_mode;
2111 best_rd = bsi->rdstat[i][mode_idx].brdcost;
2117 bsi->rdstat[i][mode_idx].brdcost = encode_inter_mb_segment(
2118 cpi, x, bsi->segment_rd - this_segment_rd, i,
2119 &bsi->rdstat[i][mode_idx].byrate, &bsi->rdstat[i][mode_idx].bdist,
2120 &bsi->rdstat[i][mode_idx].bsse, bsi->rdstat[i][mode_idx].ta,
2121 bsi->rdstat[i][mode_idx].tl, mi_row, mi_col);
2122 if (bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
2123 bsi->rdstat[i][mode_idx].brdcost +=
2124 RDCOST(x->rdmult, x->rddiv, bsi->rdstat[i][mode_idx].brate, 0);
2125 bsi->rdstat[i][mode_idx].brate += bsi->rdstat[i][mode_idx].byrate;
2126 bsi->rdstat[i][mode_idx].eobs = p->eobs[i];
2127 if (num_4x4_blocks_wide > 1)
2128 bsi->rdstat[i + 1][mode_idx].eobs = p->eobs[i + 1];
2129 if (num_4x4_blocks_high > 1)
2130 bsi->rdstat[i + 2][mode_idx].eobs = p->eobs[i + 2];
2133 if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
2134 mode_selected = this_mode;
2135 best_rd = bsi->rdstat[i][mode_idx].brdcost;
2137 } /*for each 4x4 mode*/
2139 if (best_rd == INT64_MAX) {
2141 for (iy = i + 1; iy < 4; ++iy)
2142 for (midx = 0; midx < INTER_MODES; ++midx)
2143 bsi->rdstat[iy][midx].brdcost = INT64_MAX;
2144 bsi->segment_rd = INT64_MAX;
2148 mode_idx = INTER_OFFSET(mode_selected);
2149 memcpy(t_above, bsi->rdstat[i][mode_idx].ta, sizeof(t_above));
2150 memcpy(t_left, bsi->rdstat[i][mode_idx].tl, sizeof(t_left));
2152 set_and_cost_bmi_mvs(cpi, x, xd, i, mode_selected, mode_mv[mode_selected],
2153 frame_mv, seg_mvs[i], bsi->ref_mv, x->nmvjointcost,
2156 br += bsi->rdstat[i][mode_idx].brate;
2157 bd += bsi->rdstat[i][mode_idx].bdist;
2158 block_sse += bsi->rdstat[i][mode_idx].bsse;
2159 segmentyrate += bsi->rdstat[i][mode_idx].byrate;
2160 this_segment_rd += bsi->rdstat[i][mode_idx].brdcost;
2162 if (this_segment_rd > bsi->segment_rd) {
2164 for (iy = i + 1; iy < 4; ++iy)
2165 for (midx = 0; midx < INTER_MODES; ++midx)
2166 bsi->rdstat[iy][midx].brdcost = INT64_MAX;
2167 bsi->segment_rd = INT64_MAX;
2171 } /* for each label */
2175 bsi->segment_yrate = segmentyrate;
2176 bsi->segment_rd = this_segment_rd;
2177 bsi->sse = block_sse;
2179 // update the coding decisions
2180 for (k = 0; k < 4; ++k) bsi->modes[k] = mi->bmi[k].as_mode;
2182 if (bsi->segment_rd > best_rd) return INT64_MAX;
2183 /* set it to the best */
2184 for (i = 0; i < 4; i++) {
2185 mode_idx = INTER_OFFSET(bsi->modes[i]);
2186 mi->bmi[i].as_mv[0].as_int = bsi->rdstat[i][mode_idx].mvs[0].as_int;
2187 if (has_second_ref(mi))
2188 mi->bmi[i].as_mv[1].as_int = bsi->rdstat[i][mode_idx].mvs[1].as_int;
2189 x->plane[0].eobs[i] = bsi->rdstat[i][mode_idx].eobs;
2190 mi->bmi[i].as_mode = bsi->modes[i];
2194 * used to set mbmi->mv.as_int
2196 *returntotrate = bsi->r;
2197 *returndistortion = bsi->d;
2198 *returnyrate = bsi->segment_yrate;
2199 *skippable = vp9_is_skippable_in_plane(x, BLOCK_8X8, 0);
2201 mi->mode = bsi->modes[3];
2203 return bsi->segment_rd;
2206 static void estimate_ref_frame_costs(const VP9_COMMON *cm,
2207 const MACROBLOCKD *xd, int segment_id,
2208 unsigned int *ref_costs_single,
2209 unsigned int *ref_costs_comp,
2210 vpx_prob *comp_mode_p) {
2211 int seg_ref_active =
2212 segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME);
2213 if (seg_ref_active) {
2214 memset(ref_costs_single, 0, MAX_REF_FRAMES * sizeof(*ref_costs_single));
2215 memset(ref_costs_comp, 0, MAX_REF_FRAMES * sizeof(*ref_costs_comp));
2218 vpx_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
2219 vpx_prob comp_inter_p = 128;
2221 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
2222 comp_inter_p = vp9_get_reference_mode_prob(cm, xd);
2223 *comp_mode_p = comp_inter_p;
2228 ref_costs_single[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
2230 if (cm->reference_mode != COMPOUND_REFERENCE) {
2231 vpx_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
2232 vpx_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
2233 unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
2235 if (cm->reference_mode == REFERENCE_MODE_SELECT)
2236 base_cost += vp9_cost_bit(comp_inter_p, 0);
2238 ref_costs_single[LAST_FRAME] = ref_costs_single[GOLDEN_FRAME] =
2239 ref_costs_single[ALTREF_FRAME] = base_cost;
2240 ref_costs_single[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
2241 ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
2242 ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
2243 ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
2244 ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
2246 ref_costs_single[LAST_FRAME] = 512;
2247 ref_costs_single[GOLDEN_FRAME] = 512;
2248 ref_costs_single[ALTREF_FRAME] = 512;
2250 if (cm->reference_mode != SINGLE_REFERENCE) {
2251 vpx_prob ref_comp_p = vp9_get_pred_prob_comp_ref_p(cm, xd);
2252 unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
2254 if (cm->reference_mode == REFERENCE_MODE_SELECT)
2255 base_cost += vp9_cost_bit(comp_inter_p, 1);
2257 ref_costs_comp[LAST_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 0);
2258 ref_costs_comp[GOLDEN_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 1);
2260 ref_costs_comp[LAST_FRAME] = 512;
2261 ref_costs_comp[GOLDEN_FRAME] = 512;
2266 static void store_coding_context(
2267 MACROBLOCK *x, PICK_MODE_CONTEXT *ctx, int mode_index,
2268 int64_t comp_pred_diff[REFERENCE_MODES],
2269 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS], int skippable) {
2270 MACROBLOCKD *const xd = &x->e_mbd;
2272 // Take a snapshot of the coding context so it can be
2273 // restored if we decide to encode this way
2274 ctx->skip = x->skip;
2275 ctx->skippable = skippable;
2276 ctx->best_mode_index = mode_index;
2277 ctx->mic = *xd->mi[0];
2278 ctx->mbmi_ext = *x->mbmi_ext;
2279 ctx->single_pred_diff = (int)comp_pred_diff[SINGLE_REFERENCE];
2280 ctx->comp_pred_diff = (int)comp_pred_diff[COMPOUND_REFERENCE];
2281 ctx->hybrid_pred_diff = (int)comp_pred_diff[REFERENCE_MODE_SELECT];
2283 memcpy(ctx->best_filter_diff, best_filter_diff,
2284 sizeof(*best_filter_diff) * SWITCHABLE_FILTER_CONTEXTS);
2287 static void setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
2288 MV_REFERENCE_FRAME ref_frame,
2289 BLOCK_SIZE block_size, int mi_row, int mi_col,
2290 int_mv frame_nearest_mv[MAX_REF_FRAMES],
2291 int_mv frame_near_mv[MAX_REF_FRAMES],
2292 struct buf_2d yv12_mb[4][MAX_MB_PLANE]) {
2293 const VP9_COMMON *cm = &cpi->common;
2294 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
2295 MACROBLOCKD *const xd = &x->e_mbd;
2296 MODE_INFO *const mi = xd->mi[0];
2297 int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
2298 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
2299 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2301 assert(yv12 != NULL);
2303 // TODO(jkoleszar): Is the UV buffer ever used here? If so, need to make this
2304 // use the UV scaling factors.
2305 vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
2307 // Gets an initial list of candidate vectors from neighbours and orders them
2308 vp9_find_mv_refs(cm, xd, mi, ref_frame, candidates, mi_row, mi_col,
2309 mbmi_ext->mode_context);
2311 // Candidate refinement carried out at encoder and decoder
2312 vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
2313 &frame_nearest_mv[ref_frame],
2314 &frame_near_mv[ref_frame]);
2316 // Further refinement that is encode side only to test the top few candidates
2317 // in full and choose the best as the centre point for subsequent searches.
2318 // The current implementation doesn't support scaling.
2319 if (!vp9_is_scaled(sf) && block_size >= BLOCK_8X8)
2320 vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride, ref_frame,
2324 static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
2325 int mi_row, int mi_col, int_mv *tmp_mv,
2327 MACROBLOCKD *xd = &x->e_mbd;
2328 const VP9_COMMON *cm = &cpi->common;
2329 MODE_INFO *mi = xd->mi[0];
2330 struct buf_2d backup_yv12[MAX_MB_PLANE] = { { 0, 0 } };
2331 int bestsme = INT_MAX;
2333 int sadpb = x->sadperbit16;
2335 int ref = mi->ref_frame[0];
2336 MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
2337 const MvLimits tmp_mv_limits = x->mv_limits;
2340 const YV12_BUFFER_CONFIG *scaled_ref_frame =
2341 vp9_get_scaled_ref_frame(cpi, ref);
2344 pred_mv[0] = x->mbmi_ext->ref_mvs[ref][0].as_mv;
2345 pred_mv[1] = x->mbmi_ext->ref_mvs[ref][1].as_mv;
2346 pred_mv[2] = x->pred_mv[ref];
2348 if (scaled_ref_frame) {
2350 // Swap out the reference frame for a version that's been scaled to
2351 // match the resolution of the current frame, allowing the existing
2352 // motion search code to be used without additional modifications.
2353 for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0];
2355 vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
2358 // Work out the size of the first step in the mv step search.
2359 // 0 here is maximum length first step. 1 is VPXMAX >> 1 etc.
2360 if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
2361 // Take wtd average of the step_params based on the last frame's
2362 // max mv magnitude and that based on the best ref mvs of the current
2363 // block for the given reference.
2365 (vp9_init_search_range(x->max_mv_context[ref]) + cpi->mv_step_param) /
2368 step_param = cpi->mv_step_param;
2371 if (cpi->sf.adaptive_motion_search && bsize < BLOCK_64X64) {
2373 2 * (b_width_log2_lookup[BLOCK_64X64] -
2374 VPXMIN(b_height_log2_lookup[bsize], b_width_log2_lookup[bsize]));
2375 step_param = VPXMAX(step_param, boffset);
2378 if (cpi->sf.adaptive_motion_search) {
2379 int bwl = b_width_log2_lookup[bsize];
2380 int bhl = b_height_log2_lookup[bsize];
2381 int tlevel = x->pred_mv_sad[ref] >> (bwl + bhl + 4);
2383 if (tlevel < 5) step_param += 2;
2385 // prev_mv_sad is not setup for dynamically scaled frames.
2386 if (cpi->oxcf.resize_mode != RESIZE_DYNAMIC) {
2388 for (i = LAST_FRAME; i <= ALTREF_FRAME && cm->show_frame; ++i) {
2389 if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
2390 x->pred_mv[ref].row = 0;
2391 x->pred_mv[ref].col = 0;
2392 tmp_mv->as_int = INVALID_MV;
2394 if (scaled_ref_frame) {
2396 for (i = 0; i < MAX_MB_PLANE; ++i)
2397 xd->plane[i].pre[0] = backup_yv12[i];
2405 // Note: MV limits are modified here. Always restore the original values
2406 // after full-pixel motion search.
2407 vp9_set_mv_search_range(&x->mv_limits, &ref_mv);
2409 mvp_full = pred_mv[x->mv_best_ref_index[ref]];
2414 bestsme = vp9_full_pixel_search(
2415 cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method, sadpb,
2416 cond_cost_list(cpi, cost_list), &ref_mv, &tmp_mv->as_mv, INT_MAX, 1);
2418 x->mv_limits = tmp_mv_limits;
2420 if (bestsme < INT_MAX) {
2421 uint32_t dis; /* TODO: use dis in distortion calculation later. */
2422 cpi->find_fractional_mv_step(
2423 x, &tmp_mv->as_mv, &ref_mv, cm->allow_high_precision_mv, x->errorperbit,
2424 &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
2425 cpi->sf.mv.subpel_iters_per_step, cond_cost_list(cpi, cost_list),
2426 x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref], NULL, 0, 0);
2428 *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv, x->nmvjointcost,
2429 x->mvcost, MV_COST_WEIGHT);
2431 if (cpi->sf.adaptive_motion_search) x->pred_mv[ref] = tmp_mv->as_mv;
2433 if (scaled_ref_frame) {
2435 for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
2439 static INLINE void restore_dst_buf(MACROBLOCKD *xd,
2440 uint8_t *orig_dst[MAX_MB_PLANE],
2441 int orig_dst_stride[MAX_MB_PLANE]) {
2443 for (i = 0; i < MAX_MB_PLANE; i++) {
2444 xd->plane[i].dst.buf = orig_dst[i];
2445 xd->plane[i].dst.stride = orig_dst_stride[i];
2449 // In some situations we want to discount tha pparent cost of a new motion
2450 // vector. Where there is a subtle motion field and especially where there is
2451 // low spatial complexity then it can be hard to cover the cost of a new motion
2452 // vector in a single block, even if that motion vector reduces distortion.
2453 // However, once established that vector may be usable through the nearest and
2454 // near mv modes to reduce distortion in subsequent blocks and also improve
2456 static int discount_newmv_test(const VP9_COMP *cpi, int this_mode,
2458 int_mv (*mode_mv)[MAX_REF_FRAMES],
2460 return (!cpi->rc.is_src_frame_alt_ref && (this_mode == NEWMV) &&
2461 (this_mv.as_int != 0) &&
2462 ((mode_mv[NEARESTMV][ref_frame].as_int == 0) ||
2463 (mode_mv[NEARESTMV][ref_frame].as_int == INVALID_MV)) &&
2464 ((mode_mv[NEARMV][ref_frame].as_int == 0) ||
2465 (mode_mv[NEARMV][ref_frame].as_int == INVALID_MV)));
2468 static int64_t handle_inter_mode(
2469 VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize, int *rate2,
2470 int64_t *distortion, int *skippable, int *rate_y, int *rate_uv,
2471 int *disable_skip, int_mv (*mode_mv)[MAX_REF_FRAMES], int mi_row,
2472 int mi_col, int_mv single_newmv[MAX_REF_FRAMES],
2473 INTERP_FILTER (*single_filter)[MAX_REF_FRAMES],
2474 int (*single_skippable)[MAX_REF_FRAMES], int64_t *psse,
2475 const int64_t ref_best_rd, int64_t *mask_filter, int64_t filter_cache[]) {
2476 VP9_COMMON *cm = &cpi->common;
2477 MACROBLOCKD *xd = &x->e_mbd;
2478 MODE_INFO *mi = xd->mi[0];
2479 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2480 const int is_comp_pred = has_second_ref(mi);
2481 const int this_mode = mi->mode;
2482 int_mv *frame_mv = mode_mv[this_mode];
2484 int refs[2] = { mi->ref_frame[0],
2485 (mi->ref_frame[1] < 0 ? 0 : mi->ref_frame[1]) };
2487 #if CONFIG_VP9_HIGHBITDEPTH
2488 DECLARE_ALIGNED(16, uint16_t, tmp_buf16[MAX_MB_PLANE * 64 * 64]);
2491 DECLARE_ALIGNED(16, uint8_t, tmp_buf[MAX_MB_PLANE * 64 * 64]);
2492 #endif // CONFIG_VP9_HIGHBITDEPTH
2493 int pred_exists = 0;
2495 int64_t rd, tmp_rd, best_rd = INT64_MAX;
2496 int best_needs_copy = 0;
2497 uint8_t *orig_dst[MAX_MB_PLANE];
2498 int orig_dst_stride[MAX_MB_PLANE];
2500 INTERP_FILTER best_filter = SWITCHABLE;
2501 uint8_t skip_txfm[MAX_MB_PLANE << 2] = { 0 };
2502 int64_t bsse[MAX_MB_PLANE << 2] = { 0 };
2504 int bsl = mi_width_log2_lookup[bsize];
2505 int pred_filter_search =
2506 cpi->sf.cb_pred_filter_search
2507 ? (((mi_row + mi_col) >> bsl) +
2508 get_chessboard_index(cm->current_video_frame)) &
2512 int skip_txfm_sb = 0;
2513 int64_t skip_sse_sb = INT64_MAX;
2514 int64_t distortion_y = 0, distortion_uv = 0;
2516 #if CONFIG_VP9_HIGHBITDEPTH
2517 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2518 tmp_buf = CONVERT_TO_BYTEPTR(tmp_buf16);
2520 tmp_buf = (uint8_t *)tmp_buf16;
2522 #endif // CONFIG_VP9_HIGHBITDEPTH
2524 if (pred_filter_search) {
2525 INTERP_FILTER af = SWITCHABLE, lf = SWITCHABLE;
2526 if (xd->above_mi && is_inter_block(xd->above_mi))
2527 af = xd->above_mi->interp_filter;
2528 if (xd->left_mi && is_inter_block(xd->left_mi))
2529 lf = xd->left_mi->interp_filter;
2531 if ((this_mode != NEWMV) || (af == lf)) best_filter = af;
2535 if (frame_mv[refs[0]].as_int == INVALID_MV ||
2536 frame_mv[refs[1]].as_int == INVALID_MV)
2539 if (cpi->sf.adaptive_mode_search) {
2540 if (single_filter[this_mode][refs[0]] ==
2541 single_filter[this_mode][refs[1]])
2542 best_filter = single_filter[this_mode][refs[0]];
2546 if (this_mode == NEWMV) {
2549 // Initialize mv using single prediction mode result.
2550 frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int;
2551 frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int;
2553 if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
2554 joint_motion_search(cpi, x, bsize, frame_mv, mi_row, mi_col,
2555 single_newmv, &rate_mv);
2557 rate_mv = vp9_mv_bit_cost(&frame_mv[refs[0]].as_mv,
2558 &x->mbmi_ext->ref_mvs[refs[0]][0].as_mv,
2559 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2560 rate_mv += vp9_mv_bit_cost(&frame_mv[refs[1]].as_mv,
2561 &x->mbmi_ext->ref_mvs[refs[1]][0].as_mv,
2562 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2567 single_motion_search(cpi, x, bsize, mi_row, mi_col, &tmp_mv, &rate_mv);
2568 if (tmp_mv.as_int == INVALID_MV) return INT64_MAX;
2570 frame_mv[refs[0]].as_int = xd->mi[0]->bmi[0].as_mv[0].as_int =
2572 single_newmv[refs[0]].as_int = tmp_mv.as_int;
2574 // Estimate the rate implications of a new mv but discount this
2575 // under certain circumstances where we want to help initiate a weak
2576 // motion field, where the distortion gain for a single block may not
2577 // be enough to overcome the cost of a new mv.
2578 if (discount_newmv_test(cpi, this_mode, tmp_mv, mode_mv, refs[0])) {
2579 *rate2 += VPXMAX((rate_mv / NEW_MV_DISCOUNT_FACTOR), 1);
2586 for (i = 0; i < is_comp_pred + 1; ++i) {
2587 cur_mv[i] = frame_mv[refs[i]];
2588 // Clip "next_nearest" so that it does not extend to far out of image
2589 if (this_mode != NEWMV) clamp_mv2(&cur_mv[i].as_mv, xd);
2591 if (mv_check_bounds(&x->mv_limits, &cur_mv[i].as_mv)) return INT64_MAX;
2592 mi->mv[i].as_int = cur_mv[i].as_int;
2595 // do first prediction into the destination buffer. Do the next
2596 // prediction into a temporary buffer. Then keep track of which one
2597 // of these currently holds the best predictor, and use the other
2598 // one for future predictions. In the end, copy from tmp_buf to
2599 // dst if necessary.
2600 for (i = 0; i < MAX_MB_PLANE; i++) {
2601 orig_dst[i] = xd->plane[i].dst.buf;
2602 orig_dst_stride[i] = xd->plane[i].dst.stride;
2605 // We don't include the cost of the second reference here, because there
2606 // are only two options: Last/ARF or Golden/ARF; The second one is always
2607 // known, which is ARF.
2609 // Under some circumstances we discount the cost of new mv mode to encourage
2610 // initiation of a motion field.
2611 if (discount_newmv_test(cpi, this_mode, frame_mv[refs[0]], mode_mv,
2614 VPXMIN(cost_mv_ref(cpi, this_mode, mbmi_ext->mode_context[refs[0]]),
2615 cost_mv_ref(cpi, NEARESTMV, mbmi_ext->mode_context[refs[0]]));
2617 *rate2 += cost_mv_ref(cpi, this_mode, mbmi_ext->mode_context[refs[0]]);
2620 if (RDCOST(x->rdmult, x->rddiv, *rate2, 0) > ref_best_rd &&
2621 mi->mode != NEARESTMV)
2625 // Are all MVs integer pel for Y and UV
2626 intpel_mv = !mv_has_subpel(&mi->mv[0].as_mv);
2627 if (is_comp_pred) intpel_mv &= !mv_has_subpel(&mi->mv[1].as_mv);
2629 // Search for best switchable filter by checking the variance of
2630 // pred error irrespective of whether the filter will be used
2631 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i) filter_cache[i] = INT64_MAX;
2633 if (cm->interp_filter != BILINEAR) {
2634 if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) {
2635 best_filter = EIGHTTAP;
2636 } else if (best_filter == SWITCHABLE) {
2638 int tmp_rate_sum = 0;
2639 int64_t tmp_dist_sum = 0;
2641 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
2644 int tmp_skip_sb = 0;
2645 int64_t tmp_skip_sse = INT64_MAX;
2647 mi->interp_filter = i;
2648 rs = vp9_get_switchable_rate(cpi, xd);
2649 rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
2651 if (i > 0 && intpel_mv) {
2652 rd = RDCOST(x->rdmult, x->rddiv, tmp_rate_sum, tmp_dist_sum);
2653 filter_cache[i] = rd;
2654 filter_cache[SWITCHABLE_FILTERS] =
2655 VPXMIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2656 if (cm->interp_filter == SWITCHABLE) rd += rs_rd;
2657 *mask_filter = VPXMAX(*mask_filter, rd);
2660 int64_t dist_sum = 0;
2661 if (i > 0 && cpi->sf.adaptive_interp_filter_search &&
2662 (cpi->sf.interp_filter_search_mask & (1 << i))) {
2664 dist_sum = INT64_MAX;
2668 if ((cm->interp_filter == SWITCHABLE && (!i || best_needs_copy)) ||
2669 (cm->interp_filter != SWITCHABLE &&
2670 (cm->interp_filter == mi->interp_filter ||
2671 (i == 0 && intpel_mv)))) {
2672 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2674 for (j = 0; j < MAX_MB_PLANE; j++) {
2675 xd->plane[j].dst.buf = tmp_buf + j * 64 * 64;
2676 xd->plane[j].dst.stride = 64;
2679 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2680 model_rd_for_sb(cpi, bsize, x, xd, &rate_sum, &dist_sum, &tmp_skip_sb,
2683 rd = RDCOST(x->rdmult, x->rddiv, rate_sum, dist_sum);
2684 filter_cache[i] = rd;
2685 filter_cache[SWITCHABLE_FILTERS] =
2686 VPXMIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2687 if (cm->interp_filter == SWITCHABLE) rd += rs_rd;
2688 *mask_filter = VPXMAX(*mask_filter, rd);
2690 if (i == 0 && intpel_mv) {
2691 tmp_rate_sum = rate_sum;
2692 tmp_dist_sum = dist_sum;
2696 if (i == 0 && cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2697 if (rd / 2 > ref_best_rd) {
2698 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2702 newbest = i == 0 || rd < best_rd;
2706 best_filter = mi->interp_filter;
2707 if (cm->interp_filter == SWITCHABLE && i && !intpel_mv)
2708 best_needs_copy = !best_needs_copy;
2711 if ((cm->interp_filter == SWITCHABLE && newbest) ||
2712 (cm->interp_filter != SWITCHABLE &&
2713 cm->interp_filter == mi->interp_filter)) {
2717 skip_txfm_sb = tmp_skip_sb;
2718 skip_sse_sb = tmp_skip_sse;
2719 memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2720 memcpy(bsse, x->bsse, sizeof(bsse));
2723 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2726 // Set the appropriate filter
2728 cm->interp_filter != SWITCHABLE ? cm->interp_filter : best_filter;
2729 rs = cm->interp_filter == SWITCHABLE ? vp9_get_switchable_rate(cpi, xd) : 0;
2732 if (best_needs_copy) {
2733 // again temporarily set the buffers to local memory to prevent a memcpy
2734 for (i = 0; i < MAX_MB_PLANE; i++) {
2735 xd->plane[i].dst.buf = tmp_buf + i * 64 * 64;
2736 xd->plane[i].dst.stride = 64;
2739 rd = tmp_rd + RDCOST(x->rdmult, x->rddiv, rs, 0);
2743 // Handles the special case when a filter that is not in the
2744 // switchable list (ex. bilinear) is indicated at the frame level, or
2745 // skip condition holds.
2746 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2747 model_rd_for_sb(cpi, bsize, x, xd, &tmp_rate, &tmp_dist, &skip_txfm_sb,
2749 rd = RDCOST(x->rdmult, x->rddiv, rs + tmp_rate, tmp_dist);
2750 memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2751 memcpy(bsse, x->bsse, sizeof(bsse));
2754 if (!is_comp_pred) single_filter[this_mode][refs[0]] = mi->interp_filter;
2756 if (cpi->sf.adaptive_mode_search)
2758 if (single_skippable[this_mode][refs[0]] &&
2759 single_skippable[this_mode][refs[1]])
2760 memset(skip_txfm, SKIP_TXFM_AC_DC, sizeof(skip_txfm));
2762 if (cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2763 // if current pred_error modeled rd is substantially more than the best
2764 // so far, do not bother doing full rd
2765 if (rd / 2 > ref_best_rd) {
2766 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2771 if (cm->interp_filter == SWITCHABLE) *rate2 += rs;
2773 memcpy(x->skip_txfm, skip_txfm, sizeof(skip_txfm));
2774 memcpy(x->bsse, bsse, sizeof(bsse));
2776 if (!skip_txfm_sb) {
2777 int skippable_y, skippable_uv;
2778 int64_t sseuv = INT64_MAX;
2779 int64_t rdcosty = INT64_MAX;
2781 // Y cost and distortion
2782 vp9_subtract_plane(x, bsize, 0);
2783 super_block_yrd(cpi, x, rate_y, &distortion_y, &skippable_y, psse, bsize,
2786 if (*rate_y == INT_MAX) {
2788 *distortion = INT64_MAX;
2789 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2794 *distortion += distortion_y;
2796 rdcosty = RDCOST(x->rdmult, x->rddiv, *rate2, *distortion);
2797 rdcosty = VPXMIN(rdcosty, RDCOST(x->rdmult, x->rddiv, 0, *psse));
2799 if (!super_block_uvrd(cpi, x, rate_uv, &distortion_uv, &skippable_uv,
2800 &sseuv, bsize, ref_best_rd - rdcosty)) {
2802 *distortion = INT64_MAX;
2803 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2809 *distortion += distortion_uv;
2810 *skippable = skippable_y && skippable_uv;
2815 // The cost of skip bit needs to be added.
2816 *rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2818 *distortion = skip_sse_sb;
2821 if (!is_comp_pred) single_skippable[this_mode][refs[0]] = *skippable;
2823 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2824 return 0; // The rate-distortion cost will be re-calculated by caller.
2827 void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
2828 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
2830 VP9_COMMON *const cm = &cpi->common;
2831 MACROBLOCKD *const xd = &x->e_mbd;
2832 struct macroblockd_plane *const pd = xd->plane;
2833 int rate_y = 0, rate_uv = 0, rate_y_tokenonly = 0, rate_uv_tokenonly = 0;
2834 int y_skip = 0, uv_skip = 0;
2835 int64_t dist_y = 0, dist_uv = 0;
2836 TX_SIZE max_uv_tx_size;
2839 xd->mi[0]->ref_frame[0] = INTRA_FRAME;
2840 xd->mi[0]->ref_frame[1] = NONE;
2841 // Initialize interp_filter here so we do not have to check for inter block
2842 // modes in get_pred_context_switchable_interp()
2843 xd->mi[0]->interp_filter = SWITCHABLE_FILTERS;
2845 if (bsize >= BLOCK_8X8) {
2846 if (rd_pick_intra_sby_mode(cpi, x, &rate_y, &rate_y_tokenonly, &dist_y,
2847 &y_skip, bsize, best_rd) >= best_rd) {
2848 rd_cost->rate = INT_MAX;
2853 if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate_y, &rate_y_tokenonly,
2854 &dist_y, best_rd) >= best_rd) {
2855 rd_cost->rate = INT_MAX;
2859 max_uv_tx_size = uv_txsize_lookup[bsize][xd->mi[0]->tx_size]
2860 [pd[1].subsampling_x][pd[1].subsampling_y];
2861 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly, &dist_uv,
2862 &uv_skip, VPXMAX(BLOCK_8X8, bsize), max_uv_tx_size);
2864 if (y_skip && uv_skip) {
2865 rd_cost->rate = rate_y + rate_uv - rate_y_tokenonly - rate_uv_tokenonly +
2866 vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2867 rd_cost->dist = dist_y + dist_uv;
2870 rate_y + rate_uv + vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
2871 rd_cost->dist = dist_y + dist_uv;
2874 ctx->mic = *xd->mi[0];
2875 ctx->mbmi_ext = *x->mbmi_ext;
2876 rd_cost->rdcost = RDCOST(x->rdmult, x->rddiv, rd_cost->rate, rd_cost->dist);
2879 // This function is designed to apply a bias or adjustment to an rd value based
2880 // on the relative variance of the source and reconstruction.
2881 #define VERY_LOW_VAR_THRESH 2
2882 #define LOW_VAR_THRESH 5
2883 #define VAR_MULT 100
2884 static unsigned int max_var_adjust[VP9E_CONTENT_INVALID] = { 16, 16, 100 };
2886 static void rd_variance_adjustment(VP9_COMP *cpi, MACROBLOCK *x,
2887 BLOCK_SIZE bsize, int64_t *this_rd,
2888 MV_REFERENCE_FRAME ref_frame,
2889 unsigned int source_variance) {
2890 MACROBLOCKD *const xd = &x->e_mbd;
2891 unsigned int rec_variance;
2892 unsigned int src_variance;
2893 unsigned int src_rec_min;
2894 unsigned int absvar_diff = 0;
2895 unsigned int var_factor = 0;
2896 unsigned int adj_max;
2897 vp9e_tune_content content_type = cpi->oxcf.content;
2899 if (*this_rd == INT64_MAX) return;
2901 #if CONFIG_VP9_HIGHBITDEPTH
2902 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2903 if (source_variance > 0) {
2904 rec_variance = vp9_high_get_sby_perpixel_variance(cpi, &xd->plane[0].dst,
2906 src_variance = source_variance;
2909 vp9_high_get_sby_variance(cpi, &xd->plane[0].dst, bsize, xd->bd);
2911 vp9_high_get_sby_variance(cpi, &x->plane[0].src, bsize, xd->bd);
2914 if (source_variance > 0) {
2916 vp9_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize);
2917 src_variance = source_variance;
2919 rec_variance = vp9_get_sby_variance(cpi, &xd->plane[0].dst, bsize);
2920 src_variance = vp9_get_sby_variance(cpi, &x->plane[0].src, bsize);
2924 if (source_variance > 0) {
2925 rec_variance = vp9_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize);
2926 src_variance = source_variance;
2928 rec_variance = vp9_get_sby_variance(cpi, &xd->plane[0].dst, bsize);
2929 src_variance = vp9_get_sby_variance(cpi, &x->plane[0].src, bsize);
2931 #endif // CONFIG_VP9_HIGHBITDEPTH
2933 // Lower of source (raw per pixel value) and recon variance. Note that
2934 // if the source per pixel is 0 then the recon value here will not be per
2935 // pixel (see above) so will likely be much larger.
2936 src_rec_min = VPXMIN(source_variance, rec_variance);
2938 if (src_rec_min > LOW_VAR_THRESH) return;
2940 absvar_diff = (src_variance > rec_variance) ? (src_variance - rec_variance)
2941 : (rec_variance - src_variance);
2943 adj_max = max_var_adjust[content_type];
2946 (unsigned int)((int64_t)VAR_MULT * absvar_diff) / VPXMAX(1, src_variance);
2947 var_factor = VPXMIN(adj_max, var_factor);
2949 *this_rd += (*this_rd * var_factor) / 100;
2951 if (content_type == VP9E_CONTENT_FILM) {
2952 if (src_rec_min <= VERY_LOW_VAR_THRESH) {
2953 if (ref_frame == INTRA_FRAME) *this_rd *= 2;
2954 if (bsize > 6) *this_rd *= 2;
2959 // Do we have an internal image edge (e.g. formatting bars).
2960 int vp9_internal_image_edge(VP9_COMP *cpi) {
2961 return (cpi->oxcf.pass == 2) &&
2962 ((cpi->twopass.this_frame_stats.inactive_zone_rows > 0) ||
2963 (cpi->twopass.this_frame_stats.inactive_zone_cols > 0));
2966 // Checks to see if a super block is on a horizontal image edge.
2967 // In most cases this is the "real" edge unless there are formatting
2968 // bars embedded in the stream.
2969 int vp9_active_h_edge(VP9_COMP *cpi, int mi_row, int mi_step) {
2971 int bottom_edge = cpi->common.mi_rows;
2972 int is_active_h_edge = 0;
2974 // For two pass account for any formatting bars detected.
2975 if (cpi->oxcf.pass == 2) {
2976 TWO_PASS *twopass = &cpi->twopass;
2978 // The inactive region is specified in MBs not mi units.
2979 // The image edge is in the following MB row.
2980 top_edge += (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
2982 bottom_edge -= (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
2983 bottom_edge = VPXMAX(top_edge, bottom_edge);
2986 if (((top_edge >= mi_row) && (top_edge < (mi_row + mi_step))) ||
2987 ((bottom_edge >= mi_row) && (bottom_edge < (mi_row + mi_step)))) {
2988 is_active_h_edge = 1;
2990 return is_active_h_edge;
2993 // Checks to see if a super block is on a vertical image edge.
2994 // In most cases this is the "real" edge unless there are formatting
2995 // bars embedded in the stream.
2996 int vp9_active_v_edge(VP9_COMP *cpi, int mi_col, int mi_step) {
2998 int right_edge = cpi->common.mi_cols;
2999 int is_active_v_edge = 0;
3001 // For two pass account for any formatting bars detected.
3002 if (cpi->oxcf.pass == 2) {
3003 TWO_PASS *twopass = &cpi->twopass;
3005 // The inactive region is specified in MBs not mi units.
3006 // The image edge is in the following MB row.
3007 left_edge += (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
3009 right_edge -= (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
3010 right_edge = VPXMAX(left_edge, right_edge);
3013 if (((left_edge >= mi_col) && (left_edge < (mi_col + mi_step))) ||
3014 ((right_edge >= mi_col) && (right_edge < (mi_col + mi_step)))) {
3015 is_active_v_edge = 1;
3017 return is_active_v_edge;
3020 // Checks to see if a super block is at the edge of the active image.
3021 // In most cases this is the "real" edge unless there are formatting
3022 // bars embedded in the stream.
3023 int vp9_active_edge_sb(VP9_COMP *cpi, int mi_row, int mi_col) {
3024 return vp9_active_h_edge(cpi, mi_row, MI_BLOCK_SIZE) ||
3025 vp9_active_v_edge(cpi, mi_col, MI_BLOCK_SIZE);
3028 void vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, TileDataEnc *tile_data,
3029 MACROBLOCK *x, int mi_row, int mi_col,
3030 RD_COST *rd_cost, BLOCK_SIZE bsize,
3031 PICK_MODE_CONTEXT *ctx, int64_t best_rd_so_far) {
3032 VP9_COMMON *const cm = &cpi->common;
3033 TileInfo *const tile_info = &tile_data->tile_info;
3034 RD_OPT *const rd_opt = &cpi->rd;
3035 SPEED_FEATURES *const sf = &cpi->sf;
3036 MACROBLOCKD *const xd = &x->e_mbd;
3037 MODE_INFO *const mi = xd->mi[0];
3038 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
3039 const struct segmentation *const seg = &cm->seg;
3040 PREDICTION_MODE this_mode;
3041 MV_REFERENCE_FRAME ref_frame, second_ref_frame;
3042 unsigned char segment_id = mi->segment_id;
3043 int comp_pred, i, k;
3044 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
3045 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
3046 int_mv single_newmv[MAX_REF_FRAMES] = { { 0 } };
3047 INTERP_FILTER single_inter_filter[MB_MODE_COUNT][MAX_REF_FRAMES];
3048 int single_skippable[MB_MODE_COUNT][MAX_REF_FRAMES];
3049 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
3051 int64_t best_rd = best_rd_so_far;
3052 int64_t best_pred_diff[REFERENCE_MODES];
3053 int64_t best_pred_rd[REFERENCE_MODES];
3054 int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
3055 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3056 MODE_INFO best_mbmode;
3057 int best_mode_skippable = 0;
3058 int midx, best_mode_index = -1;
3059 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3060 vpx_prob comp_mode_p;
3061 int64_t best_intra_rd = INT64_MAX;
3062 unsigned int best_pred_sse = UINT_MAX;
3063 PREDICTION_MODE best_intra_mode = DC_PRED;
3064 int rate_uv_intra[TX_SIZES], rate_uv_tokenonly[TX_SIZES];
3065 int64_t dist_uv[TX_SIZES];
3066 int skip_uv[TX_SIZES];
3067 PREDICTION_MODE mode_uv[TX_SIZES];
3068 const int intra_cost_penalty =
3069 vp9_get_intra_cost_penalty(cpi, bsize, cm->base_qindex, cm->y_dc_delta_q);
3071 uint8_t ref_frame_skip_mask[2] = { 0 };
3072 uint16_t mode_skip_mask[MAX_REF_FRAMES] = { 0 };
3073 int mode_skip_start = sf->mode_skip_start + 1;
3074 const int *const rd_threshes = rd_opt->threshes[segment_id][bsize];
3075 const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
3076 int64_t mode_threshold[MAX_MODES];
3077 int *tile_mode_map = tile_data->mode_map[bsize];
3078 int mode_map[MAX_MODES]; // Maintain mode_map information locally to avoid
3079 // lock mechanism involved with reads from
3081 const int mode_search_skip_flags = sf->mode_search_skip_flags;
3082 int64_t mask_filter = 0;
3083 int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
3085 vp9_zero(best_mbmode);
3087 x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3089 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i) filter_cache[i] = INT64_MAX;
3091 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3094 for (i = 0; i < REFERENCE_MODES; ++i) best_pred_rd[i] = INT64_MAX;
3095 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3096 best_filter_rd[i] = INT64_MAX;
3097 for (i = 0; i < TX_SIZES; i++) rate_uv_intra[i] = INT_MAX;
3098 for (i = 0; i < MAX_REF_FRAMES; ++i) x->pred_sse[i] = INT_MAX;
3099 for (i = 0; i < MB_MODE_COUNT; ++i) {
3100 for (k = 0; k < MAX_REF_FRAMES; ++k) {
3101 single_inter_filter[i][k] = SWITCHABLE;
3102 single_skippable[i][k] = 0;
3106 rd_cost->rate = INT_MAX;
3108 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3109 x->pred_mv_sad[ref_frame] = INT_MAX;
3110 if (cpi->ref_frame_flags & flag_list[ref_frame]) {
3111 assert(get_ref_frame_buffer(cpi, ref_frame) != NULL);
3112 setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
3113 frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
3115 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
3116 frame_mv[ZEROMV][ref_frame].as_int = 0;
3119 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3120 if (!(cpi->ref_frame_flags & flag_list[ref_frame])) {
3121 // Skip checking missing references in both single and compound reference
3122 // modes. Note that a mode will be skipped if both reference frames
3124 ref_frame_skip_mask[0] |= (1 << ref_frame);
3125 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3126 } else if (sf->reference_masking) {
3127 for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
3128 // Skip fixed mv modes for poor references
3129 if ((x->pred_mv_sad[ref_frame] >> 2) > x->pred_mv_sad[i]) {
3130 mode_skip_mask[ref_frame] |= INTER_NEAREST_NEAR_ZERO;
3135 // If the segment reference frame feature is enabled....
3136 // then do nothing if the current ref frame is not allowed..
3137 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
3138 get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
3139 ref_frame_skip_mask[0] |= (1 << ref_frame);
3140 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3144 // Disable this drop out case if the ref frame
3145 // segment level feature is enabled for this segment. This is to
3146 // prevent the possibility that we end up unable to pick any mode.
3147 if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
3148 // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
3149 // unless ARNR filtering is enabled in which case we want
3150 // an unfiltered alternative. We allow near/nearest as well
3151 // because they may result in zero-zero MVs but be cheaper.
3152 if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) {
3153 ref_frame_skip_mask[0] = (1 << LAST_FRAME) | (1 << GOLDEN_FRAME);
3154 ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
3155 mode_skip_mask[ALTREF_FRAME] = ~INTER_NEAREST_NEAR_ZERO;
3156 if (frame_mv[NEARMV][ALTREF_FRAME].as_int != 0)
3157 mode_skip_mask[ALTREF_FRAME] |= (1 << NEARMV);
3158 if (frame_mv[NEARESTMV][ALTREF_FRAME].as_int != 0)
3159 mode_skip_mask[ALTREF_FRAME] |= (1 << NEARESTMV);
3163 if (cpi->rc.is_src_frame_alt_ref) {
3164 if (sf->alt_ref_search_fp) {
3165 mode_skip_mask[ALTREF_FRAME] = 0;
3166 ref_frame_skip_mask[0] = ~(1 << ALTREF_FRAME);
3167 ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
3171 if (sf->alt_ref_search_fp)
3172 if (!cm->show_frame && x->pred_mv_sad[GOLDEN_FRAME] < INT_MAX)
3173 if (x->pred_mv_sad[ALTREF_FRAME] > (x->pred_mv_sad[GOLDEN_FRAME] << 1))
3174 mode_skip_mask[ALTREF_FRAME] |= INTER_ALL;
3176 if (sf->adaptive_mode_search) {
3177 if (cm->show_frame && !cpi->rc.is_src_frame_alt_ref &&
3178 cpi->rc.frames_since_golden >= 3)
3179 if (x->pred_mv_sad[GOLDEN_FRAME] > (x->pred_mv_sad[LAST_FRAME] << 1))
3180 mode_skip_mask[GOLDEN_FRAME] |= INTER_ALL;
3183 if (bsize > sf->max_intra_bsize) {
3184 ref_frame_skip_mask[0] |= (1 << INTRA_FRAME);
3185 ref_frame_skip_mask[1] |= (1 << INTRA_FRAME);
3188 mode_skip_mask[INTRA_FRAME] |=
3189 ~(sf->intra_y_mode_mask[max_txsize_lookup[bsize]]);
3191 for (i = 0; i <= LAST_NEW_MV_INDEX; ++i) mode_threshold[i] = 0;
3193 for (i = LAST_NEW_MV_INDEX + 1; i < MAX_MODES; ++i)
3194 mode_threshold[i] = ((int64_t)rd_threshes[i] * rd_thresh_freq_fact[i]) >> 5;
3196 midx = sf->schedule_mode_search ? mode_skip_start : 0;
3199 uint8_t end_pos = 0;
3200 for (i = 5; i < midx; ++i) {
3201 if (mode_threshold[tile_mode_map[i - 1]] >
3202 mode_threshold[tile_mode_map[i]]) {
3203 uint8_t tmp = tile_mode_map[i];
3204 tile_mode_map[i] = tile_mode_map[i - 1];
3205 tile_mode_map[i - 1] = tmp;
3212 memcpy(mode_map, tile_mode_map, sizeof(mode_map));
3214 for (midx = 0; midx < MAX_MODES; ++midx) {
3215 int mode_index = mode_map[midx];
3216 int mode_excluded = 0;
3217 int64_t this_rd = INT64_MAX;
3218 int disable_skip = 0;
3219 int compmode_cost = 0;
3220 int rate2 = 0, rate_y = 0, rate_uv = 0;
3221 int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
3224 int64_t total_sse = INT64_MAX;
3227 this_mode = vp9_mode_order[mode_index].mode;
3228 ref_frame = vp9_mode_order[mode_index].ref_frame[0];
3229 second_ref_frame = vp9_mode_order[mode_index].ref_frame[1];
3231 vp9_zero(x->sum_y_eobs);
3233 // Look at the reference frame of the best mode so far and set the
3234 // skip mask to look at a subset of the remaining modes.
3235 if (midx == mode_skip_start && best_mode_index >= 0) {
3236 switch (best_mbmode.ref_frame[0]) {
3237 case INTRA_FRAME: break;
3239 ref_frame_skip_mask[0] |= LAST_FRAME_MODE_MASK;
3240 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3243 ref_frame_skip_mask[0] |= GOLDEN_FRAME_MODE_MASK;
3244 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3246 case ALTREF_FRAME: ref_frame_skip_mask[0] |= ALT_REF_MODE_MASK; break;
3248 case MAX_REF_FRAMES: assert(0 && "Invalid Reference frame"); break;
3252 if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
3253 (ref_frame_skip_mask[1] & (1 << VPXMAX(0, second_ref_frame))))
3256 if (mode_skip_mask[ref_frame] & (1 << this_mode)) continue;
3258 // Test best rd so far against threshold for trying this mode.
3259 if (best_mode_skippable && sf->schedule_mode_search)
3260 mode_threshold[mode_index] <<= 1;
3262 if (best_rd < mode_threshold[mode_index]) continue;
3264 // This is only used in motion vector unit test.
3265 if (cpi->oxcf.motion_vector_unit_test && ref_frame == INTRA_FRAME) continue;
3267 if (sf->motion_field_mode_search) {
3268 const int mi_width = VPXMIN(num_8x8_blocks_wide_lookup[bsize],
3269 tile_info->mi_col_end - mi_col);
3270 const int mi_height = VPXMIN(num_8x8_blocks_high_lookup[bsize],
3271 tile_info->mi_row_end - mi_row);
3272 const int bsl = mi_width_log2_lookup[bsize];
3273 int cb_partition_search_ctrl =
3274 (((mi_row + mi_col) >> bsl) +
3275 get_chessboard_index(cm->current_video_frame)) &
3278 int const_motion = 1;
3279 int skip_ref_frame = !cb_partition_search_ctrl;
3280 MV_REFERENCE_FRAME rf = NONE;
3282 ref_mv.as_int = INVALID_MV;
3284 if ((mi_row - 1) >= tile_info->mi_row_start) {
3285 ref_mv = xd->mi[-xd->mi_stride]->mv[0];
3286 rf = xd->mi[-xd->mi_stride]->ref_frame[0];
3287 for (i = 0; i < mi_width; ++i) {
3288 ref_mi = xd->mi[-xd->mi_stride + i];
3289 const_motion &= (ref_mv.as_int == ref_mi->mv[0].as_int) &&
3290 (ref_frame == ref_mi->ref_frame[0]);
3291 skip_ref_frame &= (rf == ref_mi->ref_frame[0]);
3295 if ((mi_col - 1) >= tile_info->mi_col_start) {
3296 if (ref_mv.as_int == INVALID_MV) ref_mv = xd->mi[-1]->mv[0];
3297 if (rf == NONE) rf = xd->mi[-1]->ref_frame[0];
3298 for (i = 0; i < mi_height; ++i) {
3299 ref_mi = xd->mi[i * xd->mi_stride - 1];
3300 const_motion &= (ref_mv.as_int == ref_mi->mv[0].as_int) &&
3301 (ref_frame == ref_mi->ref_frame[0]);
3302 skip_ref_frame &= (rf == ref_mi->ref_frame[0]);
3306 if (skip_ref_frame && this_mode != NEARESTMV && this_mode != NEWMV)
3307 if (rf > INTRA_FRAME)
3308 if (ref_frame != rf) continue;
3311 if (this_mode == NEARMV || this_mode == ZEROMV) continue;
3314 comp_pred = second_ref_frame > INTRA_FRAME;
3316 if (!cpi->allow_comp_inter_inter) continue;
3318 // Skip compound inter modes if ARF is not available.
3319 if (!(cpi->ref_frame_flags & flag_list[second_ref_frame])) continue;
3321 // Do not allow compound prediction if the segment level reference frame
3322 // feature is in use as in this case there can only be one reference.
3323 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) continue;
3325 if ((mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
3326 best_mode_index >= 0 && best_mbmode.ref_frame[0] == INTRA_FRAME)
3329 mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
3331 if (ref_frame != INTRA_FRAME)
3332 mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
3335 if (ref_frame == INTRA_FRAME) {
3336 if (sf->adaptive_mode_search)
3337 if ((x->source_variance << num_pels_log2_lookup[bsize]) > best_pred_sse)
3340 if (this_mode != DC_PRED) {
3341 // Disable intra modes other than DC_PRED for blocks with low variance
3342 // Threshold for intra skipping based on source variance
3343 // TODO(debargha): Specialize the threshold for super block sizes
3344 const unsigned int skip_intra_var_thresh = 64;
3345 if ((mode_search_skip_flags & FLAG_SKIP_INTRA_LOWVAR) &&
3346 x->source_variance < skip_intra_var_thresh)
3348 // Only search the oblique modes if the best so far is
3349 // one of the neighboring directional modes
3350 if ((mode_search_skip_flags & FLAG_SKIP_INTRA_BESTINTER) &&
3351 (this_mode >= D45_PRED && this_mode <= TM_PRED)) {
3352 if (best_mode_index >= 0 && best_mbmode.ref_frame[0] > INTRA_FRAME)
3355 if (mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
3356 if (conditional_skipintra(this_mode, best_intra_mode)) continue;
3360 const MV_REFERENCE_FRAME ref_frames[2] = { ref_frame, second_ref_frame };
3361 if (!check_best_zero_mv(cpi, mbmi_ext->mode_context, frame_mv, this_mode,
3366 mi->mode = this_mode;
3367 mi->uv_mode = DC_PRED;
3368 mi->ref_frame[0] = ref_frame;
3369 mi->ref_frame[1] = second_ref_frame;
3370 // Evaluate all sub-pel filters irrespective of whether we can use
3371 // them for this frame.
3373 cm->interp_filter == SWITCHABLE ? EIGHTTAP : cm->interp_filter;
3374 mi->mv[0].as_int = mi->mv[1].as_int = 0;
3377 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
3379 // Select prediction reference frames.
3380 for (i = 0; i < MAX_MB_PLANE; i++) {
3381 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
3382 if (comp_pred) xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
3385 if (ref_frame == INTRA_FRAME) {
3387 struct macroblockd_plane *const pd = &xd->plane[1];
3388 memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
3389 super_block_yrd(cpi, x, &rate_y, &distortion_y, &skippable, NULL, bsize,
3391 if (rate_y == INT_MAX) continue;
3393 uv_tx = uv_txsize_lookup[bsize][mi->tx_size][pd->subsampling_x]
3394 [pd->subsampling_y];
3395 if (rate_uv_intra[uv_tx] == INT_MAX) {
3396 choose_intra_uv_mode(cpi, x, ctx, bsize, uv_tx, &rate_uv_intra[uv_tx],
3397 &rate_uv_tokenonly[uv_tx], &dist_uv[uv_tx],
3398 &skip_uv[uv_tx], &mode_uv[uv_tx]);
3401 rate_uv = rate_uv_tokenonly[uv_tx];
3402 distortion_uv = dist_uv[uv_tx];
3403 skippable = skippable && skip_uv[uv_tx];
3404 mi->uv_mode = mode_uv[uv_tx];
3406 rate2 = rate_y + cpi->mbmode_cost[mi->mode] + rate_uv_intra[uv_tx];
3407 if (this_mode != DC_PRED && this_mode != TM_PRED)
3408 rate2 += intra_cost_penalty;
3409 distortion2 = distortion_y + distortion_uv;
3411 this_rd = handle_inter_mode(
3412 cpi, x, bsize, &rate2, &distortion2, &skippable, &rate_y, &rate_uv,
3413 &disable_skip, frame_mv, mi_row, mi_col, single_newmv,
3414 single_inter_filter, single_skippable, &total_sse, best_rd,
3415 &mask_filter, filter_cache);
3416 if (this_rd == INT64_MAX) continue;
3418 compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
3420 if (cm->reference_mode == REFERENCE_MODE_SELECT) rate2 += compmode_cost;
3423 // Estimate the reference frame signaling cost and add it
3424 // to the rolling cost variable.
3426 rate2 += ref_costs_comp[ref_frame];
3428 rate2 += ref_costs_single[ref_frame];
3431 if (!disable_skip) {
3432 const vpx_prob skip_prob = vp9_get_skip_prob(cm, xd);
3433 const int skip_cost0 = vp9_cost_bit(skip_prob, 0);
3434 const int skip_cost1 = vp9_cost_bit(skip_prob, 1);
3437 // Back out the coefficient coding costs
3438 rate2 -= (rate_y + rate_uv);
3440 // Cost the skip mb case
3441 rate2 += skip_cost1;
3442 } else if (ref_frame != INTRA_FRAME && !xd->lossless) {
3443 if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv + skip_cost0,
3445 RDCOST(x->rdmult, x->rddiv, skip_cost1, total_sse)) {
3446 // Add in the cost of the no skip flag.
3447 rate2 += skip_cost0;
3449 // FIXME(rbultje) make this work for splitmv also
3450 assert(total_sse >= 0);
3452 rate2 += skip_cost1;
3453 distortion2 = total_sse;
3454 rate2 -= (rate_y + rate_uv);
3458 // Add in the cost of the no skip flag.
3459 rate2 += skip_cost0;
3462 // Calculate the final RD estimate for this mode.
3463 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3466 // Apply an adjustment to the rd value based on the similarity of the
3467 // source variance and reconstructed variance.
3468 rd_variance_adjustment(cpi, x, bsize, &this_rd, ref_frame,
3469 x->source_variance);
3471 if (ref_frame == INTRA_FRAME) {
3472 // Keep record of best intra rd
3473 if (this_rd < best_intra_rd) {
3474 best_intra_rd = this_rd;
3475 best_intra_mode = mi->mode;
3479 if (!disable_skip && ref_frame == INTRA_FRAME) {
3480 for (i = 0; i < REFERENCE_MODES; ++i)
3481 best_pred_rd[i] = VPXMIN(best_pred_rd[i], this_rd);
3482 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3483 best_filter_rd[i] = VPXMIN(best_filter_rd[i], this_rd);
3486 // Did this mode help.. i.e. is it the new best mode
3487 if (this_rd < best_rd || x->skip) {
3488 int max_plane = MAX_MB_PLANE;
3489 if (!mode_excluded) {
3490 // Note index of best mode so far
3491 best_mode_index = mode_index;
3493 if (ref_frame == INTRA_FRAME) {
3494 /* required for left and above block mv */
3495 mi->mv[0].as_int = 0;
3497 // Initialize interp_filter here so we do not have to check for
3498 // inter block modes in get_pred_context_switchable_interp()
3499 mi->interp_filter = SWITCHABLE_FILTERS;
3501 best_pred_sse = x->pred_sse[ref_frame];
3504 rd_cost->rate = rate2;
3505 rd_cost->dist = distortion2;
3506 rd_cost->rdcost = this_rd;
3509 best_skip2 = this_skip2;
3510 best_mode_skippable = skippable;
3512 if (!x->select_tx_size) swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
3513 memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mi->tx_size],
3514 sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
3515 ctx->sum_y_eobs = x->sum_y_eobs[mi->tx_size];
3517 // TODO(debargha): enhance this test with a better distortion prediction
3518 // based on qp, activity mask and history
3519 if ((mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
3520 (mode_index > MIN_EARLY_TERM_INDEX)) {
3521 int qstep = xd->plane[0].dequant[1];
3522 // TODO(debargha): Enhance this by specializing for each mode_index
3524 #if CONFIG_VP9_HIGHBITDEPTH
3525 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
3526 qstep >>= (xd->bd - 8);
3528 #endif // CONFIG_VP9_HIGHBITDEPTH
3529 if (x->source_variance < UINT_MAX) {
3530 const int var_adjust = (x->source_variance < 16);
3531 scale -= var_adjust;
3533 if (ref_frame > INTRA_FRAME && distortion2 * scale < qstep * qstep) {
3540 /* keep record of best compound/single-only prediction */
3541 if (!disable_skip && ref_frame != INTRA_FRAME) {
3542 int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
3544 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
3545 single_rate = rate2 - compmode_cost;
3546 hybrid_rate = rate2;
3548 single_rate = rate2;
3549 hybrid_rate = rate2 + compmode_cost;
3552 single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
3553 hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
3556 if (single_rd < best_pred_rd[SINGLE_REFERENCE])
3557 best_pred_rd[SINGLE_REFERENCE] = single_rd;
3559 if (single_rd < best_pred_rd[COMPOUND_REFERENCE])
3560 best_pred_rd[COMPOUND_REFERENCE] = single_rd;
3562 if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
3563 best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
3565 /* keep record of best filter type */
3566 if (!mode_excluded && cm->interp_filter != BILINEAR) {
3568 filter_cache[cm->interp_filter == SWITCHABLE ? SWITCHABLE_FILTERS
3569 : cm->interp_filter];
3571 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3573 if (ref == INT64_MAX)
3575 else if (filter_cache[i] == INT64_MAX)
3576 // when early termination is triggered, the encoder does not have
3577 // access to the rate-distortion cost. it only knows that the cost
3578 // should be above the maximum valid value. hence it takes the known
3579 // maximum plus an arbitrary constant as the rate-distortion cost.
3580 adj_rd = mask_filter - ref + 10;
3582 adj_rd = filter_cache[i] - ref;
3585 best_filter_rd[i] = VPXMIN(best_filter_rd[i], adj_rd);
3590 if (early_term) break;
3592 if (x->skip && !comp_pred) break;
3595 // The inter modes' rate costs are not calculated precisely in some cases.
3596 // Therefore, sometimes, NEWMV is chosen instead of NEARESTMV, NEARMV, and
3597 // ZEROMV. Here, checks are added for those cases, and the mode decisions
3599 if (best_mbmode.mode == NEWMV) {
3600 const MV_REFERENCE_FRAME refs[2] = { best_mbmode.ref_frame[0],
3601 best_mbmode.ref_frame[1] };
3602 int comp_pred_mode = refs[1] > INTRA_FRAME;
3604 if (frame_mv[NEARESTMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3606 frame_mv[NEARESTMV][refs[1]].as_int == best_mbmode.mv[1].as_int) ||
3608 best_mbmode.mode = NEARESTMV;
3609 else if (frame_mv[NEARMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3611 frame_mv[NEARMV][refs[1]].as_int == best_mbmode.mv[1].as_int) ||
3613 best_mbmode.mode = NEARMV;
3614 else if (best_mbmode.mv[0].as_int == 0 &&
3615 ((comp_pred_mode && best_mbmode.mv[1].as_int == 0) ||
3617 best_mbmode.mode = ZEROMV;
3620 if (best_mode_index < 0 || best_rd >= best_rd_so_far) {
3621 // If adaptive interp filter is enabled, then the current leaf node of 8x8
3622 // data is needed for sub8x8. Hence preserve the context.
3623 if (cpi->row_mt && bsize == BLOCK_8X8) ctx->mic = *xd->mi[0];
3624 rd_cost->rate = INT_MAX;
3625 rd_cost->rdcost = INT64_MAX;
3629 // If we used an estimate for the uv intra rd in the loop above...
3630 if (sf->use_uv_intra_rd_estimate) {
3631 // Do Intra UV best rd mode selection if best mode choice above was intra.
3632 if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
3635 uv_tx_size = get_uv_tx_size(mi, &xd->plane[1]);
3636 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra[uv_tx_size],
3637 &rate_uv_tokenonly[uv_tx_size],
3638 &dist_uv[uv_tx_size], &skip_uv[uv_tx_size],
3639 bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize,
3644 assert((cm->interp_filter == SWITCHABLE) ||
3645 (cm->interp_filter == best_mbmode.interp_filter) ||
3646 !is_inter_block(&best_mbmode));
3648 if (!cpi->rc.is_src_frame_alt_ref)
3649 vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
3650 sf->adaptive_rd_thresh, bsize, best_mode_index);
3654 x->skip |= best_skip2;
3656 for (i = 0; i < REFERENCE_MODES; ++i) {
3657 if (best_pred_rd[i] == INT64_MAX)
3658 best_pred_diff[i] = INT_MIN;
3660 best_pred_diff[i] = best_rd - best_pred_rd[i];
3664 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3665 if (best_filter_rd[i] == INT64_MAX)
3666 best_filter_diff[i] = 0;
3668 best_filter_diff[i] = best_rd - best_filter_rd[i];
3670 if (cm->interp_filter == SWITCHABLE)
3671 assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
3673 vp9_zero(best_filter_diff);
3676 // TODO(yunqingwang): Moving this line in front of the above best_filter_diff
3677 // updating code causes PSNR loss. Need to figure out the confliction.
3678 x->skip |= best_mode_skippable;
3680 if (!x->skip && !x->select_tx_size) {
3681 int has_high_freq_coeff = 0;
3683 int max_plane = is_inter_block(xd->mi[0]) ? MAX_MB_PLANE : 1;
3684 for (plane = 0; plane < max_plane; ++plane) {
3685 x->plane[plane].eobs = ctx->eobs_pbuf[plane][1];
3686 has_high_freq_coeff |= vp9_has_high_freq_in_plane(x, bsize, plane);
3689 for (plane = max_plane; plane < MAX_MB_PLANE; ++plane) {
3690 x->plane[plane].eobs = ctx->eobs_pbuf[plane][2];
3691 has_high_freq_coeff |= vp9_has_high_freq_in_plane(x, bsize, plane);
3694 best_mode_skippable |= !has_high_freq_coeff;
3697 assert(best_mode_index >= 0);
3699 store_coding_context(x, ctx, best_mode_index, best_pred_diff,
3700 best_filter_diff, best_mode_skippable);
3703 void vp9_rd_pick_inter_mode_sb_seg_skip(VP9_COMP *cpi, TileDataEnc *tile_data,
3704 MACROBLOCK *x, RD_COST *rd_cost,
3706 PICK_MODE_CONTEXT *ctx,
3707 int64_t best_rd_so_far) {
3708 VP9_COMMON *const cm = &cpi->common;
3709 MACROBLOCKD *const xd = &x->e_mbd;
3710 MODE_INFO *const mi = xd->mi[0];
3711 unsigned char segment_id = mi->segment_id;
3712 const int comp_pred = 0;
3714 int64_t best_pred_diff[REFERENCE_MODES];
3715 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3716 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3717 vpx_prob comp_mode_p;
3718 INTERP_FILTER best_filter = SWITCHABLE;
3719 int64_t this_rd = INT64_MAX;
3721 const int64_t distortion2 = 0;
3723 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3725 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3728 for (i = 0; i < MAX_REF_FRAMES; ++i) x->pred_sse[i] = INT_MAX;
3729 for (i = LAST_FRAME; i < MAX_REF_FRAMES; ++i) x->pred_mv_sad[i] = INT_MAX;
3731 rd_cost->rate = INT_MAX;
3733 assert(segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP));
3736 mi->uv_mode = DC_PRED;
3737 mi->ref_frame[0] = LAST_FRAME;
3738 mi->ref_frame[1] = NONE;
3739 mi->mv[0].as_int = 0;
3742 ctx->sum_y_eobs = 0;
3744 if (cm->interp_filter != BILINEAR) {
3745 best_filter = EIGHTTAP;
3746 if (cm->interp_filter == SWITCHABLE &&
3747 x->source_variance >= cpi->sf.disable_filter_search_var_thresh) {
3749 int best_rs = INT_MAX;
3750 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
3751 mi->interp_filter = i;
3752 rs = vp9_get_switchable_rate(cpi, xd);
3755 best_filter = mi->interp_filter;
3760 // Set the appropriate filter
3761 if (cm->interp_filter == SWITCHABLE) {
3762 mi->interp_filter = best_filter;
3763 rate2 += vp9_get_switchable_rate(cpi, xd);
3765 mi->interp_filter = cm->interp_filter;
3768 if (cm->reference_mode == REFERENCE_MODE_SELECT)
3769 rate2 += vp9_cost_bit(comp_mode_p, comp_pred);
3771 // Estimate the reference frame signaling cost and add it
3772 // to the rolling cost variable.
3773 rate2 += ref_costs_single[LAST_FRAME];
3774 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3776 rd_cost->rate = rate2;
3777 rd_cost->dist = distortion2;
3778 rd_cost->rdcost = this_rd;
3780 if (this_rd >= best_rd_so_far) {
3781 rd_cost->rate = INT_MAX;
3782 rd_cost->rdcost = INT64_MAX;
3786 assert((cm->interp_filter == SWITCHABLE) ||
3787 (cm->interp_filter == mi->interp_filter));
3789 vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
3790 cpi->sf.adaptive_rd_thresh, bsize, THR_ZEROMV);
3792 vp9_zero(best_pred_diff);
3793 vp9_zero(best_filter_diff);
3795 if (!x->select_tx_size) swap_block_ptr(x, ctx, 1, 0, 0, MAX_MB_PLANE);
3796 store_coding_context(x, ctx, THR_ZEROMV, best_pred_diff, best_filter_diff, 0);
3799 void vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, TileDataEnc *tile_data,
3800 MACROBLOCK *x, int mi_row, int mi_col,
3801 RD_COST *rd_cost, BLOCK_SIZE bsize,
3802 PICK_MODE_CONTEXT *ctx,
3803 int64_t best_rd_so_far) {
3804 VP9_COMMON *const cm = &cpi->common;
3805 RD_OPT *const rd_opt = &cpi->rd;
3806 SPEED_FEATURES *const sf = &cpi->sf;
3807 MACROBLOCKD *const xd = &x->e_mbd;
3808 MODE_INFO *const mi = xd->mi[0];
3809 const struct segmentation *const seg = &cm->seg;
3810 MV_REFERENCE_FRAME ref_frame, second_ref_frame;
3811 unsigned char segment_id = mi->segment_id;
3813 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
3814 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
3815 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
3817 int64_t best_rd = best_rd_so_far;
3818 int64_t best_yrd = best_rd_so_far; // FIXME(rbultje) more precise
3819 int64_t best_pred_diff[REFERENCE_MODES];
3820 int64_t best_pred_rd[REFERENCE_MODES];
3821 int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
3822 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3823 MODE_INFO best_mbmode;
3824 int ref_index, best_ref_index = 0;
3825 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3826 vpx_prob comp_mode_p;
3827 INTERP_FILTER tmp_best_filter = SWITCHABLE;
3828 int rate_uv_intra, rate_uv_tokenonly;
3831 PREDICTION_MODE mode_uv = DC_PRED;
3832 const int intra_cost_penalty =
3833 vp9_get_intra_cost_penalty(cpi, bsize, cm->base_qindex, cm->y_dc_delta_q);
3834 int_mv seg_mvs[4][MAX_REF_FRAMES];
3835 b_mode_info best_bmodes[4];
3837 int ref_frame_skip_mask[2] = { 0 };
3838 int64_t mask_filter = 0;
3839 int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
3840 int internal_active_edge =
3841 vp9_active_edge_sb(cpi, mi_row, mi_col) && vp9_internal_image_edge(cpi);
3842 const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
3844 x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3845 memset(x->zcoeff_blk[TX_4X4], 0, 4);
3846 vp9_zero(best_mbmode);
3848 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i) filter_cache[i] = INT64_MAX;
3850 for (i = 0; i < 4; i++) {
3852 for (j = 0; j < MAX_REF_FRAMES; j++) seg_mvs[i][j].as_int = INVALID_MV;
3855 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3858 for (i = 0; i < REFERENCE_MODES; ++i) best_pred_rd[i] = INT64_MAX;
3859 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3860 best_filter_rd[i] = INT64_MAX;
3861 rate_uv_intra = INT_MAX;
3863 rd_cost->rate = INT_MAX;
3865 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ref_frame++) {
3866 if (cpi->ref_frame_flags & flag_list[ref_frame]) {
3867 setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
3868 frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
3870 ref_frame_skip_mask[0] |= (1 << ref_frame);
3871 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3873 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
3874 frame_mv[ZEROMV][ref_frame].as_int = 0;
3877 for (ref_index = 0; ref_index < MAX_REFS; ++ref_index) {
3878 int mode_excluded = 0;
3879 int64_t this_rd = INT64_MAX;
3880 int disable_skip = 0;
3881 int compmode_cost = 0;
3882 int rate2 = 0, rate_y = 0, rate_uv = 0;
3883 int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
3887 int64_t total_sse = INT_MAX;
3889 struct buf_2d backup_yv12[2][MAX_MB_PLANE];
3891 ref_frame = vp9_ref_order[ref_index].ref_frame[0];
3892 second_ref_frame = vp9_ref_order[ref_index].ref_frame[1];
3894 vp9_zero(x->sum_y_eobs);
3896 #if CONFIG_BETTER_HW_COMPATIBILITY
3897 // forbid 8X4 and 4X8 partitions if any reference frame is scaled.
3898 if (bsize == BLOCK_8X4 || bsize == BLOCK_4X8) {
3899 int ref_scaled = vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf);
3900 if (second_ref_frame > INTRA_FRAME)
3901 ref_scaled += vp9_is_scaled(&cm->frame_refs[second_ref_frame - 1].sf);
3902 if (ref_scaled) continue;
3905 // Look at the reference frame of the best mode so far and set the
3906 // skip mask to look at a subset of the remaining modes.
3907 if (ref_index > 2 && sf->mode_skip_start < MAX_MODES) {
3908 if (ref_index == 3) {
3909 switch (best_mbmode.ref_frame[0]) {
3910 case INTRA_FRAME: break;
3912 ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME);
3913 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3916 ref_frame_skip_mask[0] |= (1 << LAST_FRAME) | (1 << ALTREF_FRAME);
3917 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3920 ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << LAST_FRAME);
3923 case MAX_REF_FRAMES: assert(0 && "Invalid Reference frame"); break;
3928 if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
3929 (ref_frame_skip_mask[1] & (1 << VPXMAX(0, second_ref_frame))))
3932 // Test best rd so far against threshold for trying this mode.
3933 if (!internal_active_edge &&
3934 rd_less_than_thresh(best_rd,
3935 rd_opt->threshes[segment_id][bsize][ref_index],
3936 &rd_thresh_freq_fact[ref_index]))
3939 // This is only used in motion vector unit test.
3940 if (cpi->oxcf.motion_vector_unit_test && ref_frame == INTRA_FRAME) continue;
3942 comp_pred = second_ref_frame > INTRA_FRAME;
3944 if (!cpi->allow_comp_inter_inter) continue;
3945 if (!(cpi->ref_frame_flags & flag_list[second_ref_frame])) continue;
3946 // Do not allow compound prediction if the segment level reference frame
3947 // feature is in use as in this case there can only be one reference.
3948 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) continue;
3950 if ((sf->mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
3951 best_mbmode.ref_frame[0] == INTRA_FRAME)
3956 mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
3957 else if (ref_frame != INTRA_FRAME)
3958 mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
3960 // If the segment reference frame feature is enabled....
3961 // then do nothing if the current ref frame is not allowed..
3962 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
3963 get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
3965 // Disable this drop out case if the ref frame
3966 // segment level feature is enabled for this segment. This is to
3967 // prevent the possibility that we end up unable to pick any mode.
3968 } else if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
3969 // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
3970 // unless ARNR filtering is enabled in which case we want
3971 // an unfiltered alternative. We allow near/nearest as well
3972 // because they may result in zero-zero MVs but be cheaper.
3973 if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0))
3977 mi->tx_size = TX_4X4;
3978 mi->uv_mode = DC_PRED;
3979 mi->ref_frame[0] = ref_frame;
3980 mi->ref_frame[1] = second_ref_frame;
3981 // Evaluate all sub-pel filters irrespective of whether we can use
3982 // them for this frame.
3984 cm->interp_filter == SWITCHABLE ? EIGHTTAP : cm->interp_filter;
3986 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
3988 // Select prediction reference frames.
3989 for (i = 0; i < MAX_MB_PLANE; i++) {
3990 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
3991 if (comp_pred) xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
3994 if (ref_frame == INTRA_FRAME) {
3996 if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate, &rate_y, &distortion_y,
3997 best_rd) >= best_rd)
4000 rate2 += intra_cost_penalty;
4001 distortion2 += distortion_y;
4003 if (rate_uv_intra == INT_MAX) {
4004 choose_intra_uv_mode(cpi, x, ctx, bsize, TX_4X4, &rate_uv_intra,
4005 &rate_uv_tokenonly, &dist_uv, &skip_uv, &mode_uv);
4007 rate2 += rate_uv_intra;
4008 rate_uv = rate_uv_tokenonly;
4009 distortion2 += dist_uv;
4010 distortion_uv = dist_uv;
4011 mi->uv_mode = mode_uv;
4015 int64_t this_rd_thresh;
4016 int64_t tmp_rd, tmp_best_rd = INT64_MAX, tmp_best_rdu = INT64_MAX;
4017 int tmp_best_rate = INT_MAX, tmp_best_ratey = INT_MAX;
4018 int64_t tmp_best_distortion = INT_MAX, tmp_best_sse, uv_sse;
4019 int tmp_best_skippable = 0;
4020 int switchable_filter_index;
4021 int_mv *second_ref =
4022 comp_pred ? &x->mbmi_ext->ref_mvs[second_ref_frame][0] : NULL;
4023 b_mode_info tmp_best_bmodes[16];
4024 MODE_INFO tmp_best_mbmode;
4025 BEST_SEG_INFO bsi[SWITCHABLE_FILTERS];
4026 int pred_exists = 0;
4029 YV12_BUFFER_CONFIG *scaled_ref_frame[2] = { NULL, NULL };
4032 for (ref = 0; ref < 2; ++ref) {
4033 scaled_ref_frame[ref] =
4034 mi->ref_frame[ref] > INTRA_FRAME
4035 ? vp9_get_scaled_ref_frame(cpi, mi->ref_frame[ref])
4038 if (scaled_ref_frame[ref]) {
4040 // Swap out the reference frame for a version that's been scaled to
4041 // match the resolution of the current frame, allowing the existing
4042 // motion search code to be used without additional modifications.
4043 for (i = 0; i < MAX_MB_PLANE; i++)
4044 backup_yv12[ref][i] = xd->plane[i].pre[ref];
4045 vp9_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
4050 this_rd_thresh = (ref_frame == LAST_FRAME)
4051 ? rd_opt->threshes[segment_id][bsize][THR_LAST]
4052 : rd_opt->threshes[segment_id][bsize][THR_ALTR];
4053 this_rd_thresh = (ref_frame == GOLDEN_FRAME)
4054 ? rd_opt->threshes[segment_id][bsize][THR_GOLD]
4056 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
4057 filter_cache[i] = INT64_MAX;
4059 if (cm->interp_filter != BILINEAR) {
4060 tmp_best_filter = EIGHTTAP;
4061 if (x->source_variance < sf->disable_filter_search_var_thresh) {
4062 tmp_best_filter = EIGHTTAP;
4063 } else if (sf->adaptive_pred_interp_filter == 1 &&
4064 ctx->pred_interp_filter < SWITCHABLE) {
4065 tmp_best_filter = ctx->pred_interp_filter;
4066 } else if (sf->adaptive_pred_interp_filter == 2) {
4067 tmp_best_filter = ctx->pred_interp_filter < SWITCHABLE
4068 ? ctx->pred_interp_filter
4071 for (switchable_filter_index = 0;
4072 switchable_filter_index < SWITCHABLE_FILTERS;
4073 ++switchable_filter_index) {
4076 MB_MODE_INFO_EXT *mbmi_ext = x->mbmi_ext;
4077 mi->interp_filter = switchable_filter_index;
4078 tmp_rd = rd_pick_best_sub8x8_mode(
4079 cpi, x, &mbmi_ext->ref_mvs[ref_frame][0], second_ref, best_yrd,
4080 &rate, &rate_y, &distortion, &skippable, &total_sse,
4081 (int)this_rd_thresh, seg_mvs, bsi, switchable_filter_index,
4084 if (tmp_rd == INT64_MAX) continue;
4085 rs = vp9_get_switchable_rate(cpi, xd);
4086 rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
4087 filter_cache[switchable_filter_index] = tmp_rd;
4088 filter_cache[SWITCHABLE_FILTERS] =
4089 VPXMIN(filter_cache[SWITCHABLE_FILTERS], tmp_rd + rs_rd);
4090 if (cm->interp_filter == SWITCHABLE) tmp_rd += rs_rd;
4092 mask_filter = VPXMAX(mask_filter, tmp_rd);
4094 newbest = (tmp_rd < tmp_best_rd);
4096 tmp_best_filter = mi->interp_filter;
4097 tmp_best_rd = tmp_rd;
4099 if ((newbest && cm->interp_filter == SWITCHABLE) ||
4100 (mi->interp_filter == cm->interp_filter &&
4101 cm->interp_filter != SWITCHABLE)) {
4102 tmp_best_rdu = tmp_rd;
4103 tmp_best_rate = rate;
4104 tmp_best_ratey = rate_y;
4105 tmp_best_distortion = distortion;
4106 tmp_best_sse = total_sse;
4107 tmp_best_skippable = skippable;
4108 tmp_best_mbmode = *mi;
4109 for (i = 0; i < 4; i++) {
4110 tmp_best_bmodes[i] = xd->mi[0]->bmi[i];
4111 x->zcoeff_blk[TX_4X4][i] = !x->plane[0].eobs[i];
4112 x->sum_y_eobs[TX_4X4] += x->plane[0].eobs[i];
4115 if (switchable_filter_index == 0 && sf->use_rd_breakout &&
4116 best_rd < INT64_MAX) {
4117 if (tmp_best_rdu / 2 > best_rd) {
4118 // skip searching the other filters if the first is
4119 // already substantially larger than the best so far
4120 tmp_best_filter = mi->interp_filter;
4121 tmp_best_rdu = INT64_MAX;
4126 } // switchable_filter_index loop
4130 if (tmp_best_rdu == INT64_MAX && pred_exists) continue;
4132 mi->interp_filter = (cm->interp_filter == SWITCHABLE ? tmp_best_filter
4133 : cm->interp_filter);
4135 // Handles the special case when a filter that is not in the
4136 // switchable list (bilinear, 6-tap) is indicated at the frame level
4137 tmp_rd = rd_pick_best_sub8x8_mode(
4138 cpi, x, &x->mbmi_ext->ref_mvs[ref_frame][0], second_ref, best_yrd,
4139 &rate, &rate_y, &distortion, &skippable, &total_sse,
4140 (int)this_rd_thresh, seg_mvs, bsi, 0, mi_row, mi_col);
4141 if (tmp_rd == INT64_MAX) continue;
4143 total_sse = tmp_best_sse;
4144 rate = tmp_best_rate;
4145 rate_y = tmp_best_ratey;
4146 distortion = tmp_best_distortion;
4147 skippable = tmp_best_skippable;
4148 *mi = tmp_best_mbmode;
4149 for (i = 0; i < 4; i++) xd->mi[0]->bmi[i] = tmp_best_bmodes[i];
4153 distortion2 += distortion;
4155 if (cm->interp_filter == SWITCHABLE)
4156 rate2 += vp9_get_switchable_rate(cpi, xd);
4159 mode_excluded = comp_pred ? cm->reference_mode == SINGLE_REFERENCE
4160 : cm->reference_mode == COMPOUND_REFERENCE;
4162 compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
4165 best_rd - VPXMIN(RDCOST(x->rdmult, x->rddiv, rate2, distortion2),
4166 RDCOST(x->rdmult, x->rddiv, 0, total_sse));
4168 if (tmp_best_rdu > 0) {
4169 // If even the 'Y' rd value of split is higher than best so far
4170 // then dont bother looking at UV
4171 vp9_build_inter_predictors_sbuv(&x->e_mbd, mi_row, mi_col, BLOCK_8X8);
4172 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
4173 if (!super_block_uvrd(cpi, x, &rate_uv, &distortion_uv, &uv_skippable,
4174 &uv_sse, BLOCK_8X8, tmp_best_rdu)) {
4175 for (ref = 0; ref < 2; ++ref) {
4176 if (scaled_ref_frame[ref]) {
4178 for (i = 0; i < MAX_MB_PLANE; ++i)
4179 xd->plane[i].pre[ref] = backup_yv12[ref][i];
4186 distortion2 += distortion_uv;
4187 skippable = skippable && uv_skippable;
4188 total_sse += uv_sse;
4191 for (ref = 0; ref < 2; ++ref) {
4192 if (scaled_ref_frame[ref]) {
4193 // Restore the prediction frame pointers to their unscaled versions.
4195 for (i = 0; i < MAX_MB_PLANE; ++i)
4196 xd->plane[i].pre[ref] = backup_yv12[ref][i];
4201 if (cm->reference_mode == REFERENCE_MODE_SELECT) rate2 += compmode_cost;
4203 // Estimate the reference frame signaling cost and add it
4204 // to the rolling cost variable.
4205 if (second_ref_frame > INTRA_FRAME) {
4206 rate2 += ref_costs_comp[ref_frame];
4208 rate2 += ref_costs_single[ref_frame];
4211 if (!disable_skip) {
4212 const vpx_prob skip_prob = vp9_get_skip_prob(cm, xd);
4213 const int skip_cost0 = vp9_cost_bit(skip_prob, 0);
4214 const int skip_cost1 = vp9_cost_bit(skip_prob, 1);
4216 // Skip is never coded at the segment level for sub8x8 blocks and instead
4217 // always coded in the bitstream at the mode info level.
4218 if (ref_frame != INTRA_FRAME && !xd->lossless) {
4219 if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv + skip_cost0,
4221 RDCOST(x->rdmult, x->rddiv, skip_cost1, total_sse)) {
4222 // Add in the cost of the no skip flag.
4223 rate2 += skip_cost0;
4225 // FIXME(rbultje) make this work for splitmv also
4226 rate2 += skip_cost1;
4227 distortion2 = total_sse;
4228 assert(total_sse >= 0);
4229 rate2 -= (rate_y + rate_uv);
4235 // Add in the cost of the no skip flag.
4236 rate2 += skip_cost0;
4239 // Calculate the final RD estimate for this mode.
4240 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
4243 if (!disable_skip && ref_frame == INTRA_FRAME) {
4244 for (i = 0; i < REFERENCE_MODES; ++i)
4245 best_pred_rd[i] = VPXMIN(best_pred_rd[i], this_rd);
4246 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
4247 best_filter_rd[i] = VPXMIN(best_filter_rd[i], this_rd);
4250 // Did this mode help.. i.e. is it the new best mode
4251 if (this_rd < best_rd || x->skip) {
4252 if (!mode_excluded) {
4253 int max_plane = MAX_MB_PLANE;
4254 // Note index of best mode so far
4255 best_ref_index = ref_index;
4257 if (ref_frame == INTRA_FRAME) {
4258 /* required for left and above block mv */
4259 mi->mv[0].as_int = 0;
4261 // Initialize interp_filter here so we do not have to check for
4262 // inter block modes in get_pred_context_switchable_interp()
4263 mi->interp_filter = SWITCHABLE_FILTERS;
4266 rd_cost->rate = rate2;
4267 rd_cost->dist = distortion2;
4268 rd_cost->rdcost = this_rd;
4271 best_rd - RDCOST(x->rdmult, x->rddiv, rate_uv, distortion_uv);
4273 best_skip2 = this_skip2;
4274 if (!x->select_tx_size) swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
4275 memcpy(ctx->zcoeff_blk, x->zcoeff_blk[TX_4X4],
4276 sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
4277 ctx->sum_y_eobs = x->sum_y_eobs[TX_4X4];
4279 for (i = 0; i < 4; i++) best_bmodes[i] = xd->mi[0]->bmi[i];
4281 // TODO(debargha): enhance this test with a better distortion prediction
4282 // based on qp, activity mask and history
4283 if ((sf->mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
4284 (ref_index > MIN_EARLY_TERM_INDEX)) {
4285 int qstep = xd->plane[0].dequant[1];
4286 // TODO(debargha): Enhance this by specializing for each mode_index
4288 #if CONFIG_VP9_HIGHBITDEPTH
4289 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
4290 qstep >>= (xd->bd - 8);
4292 #endif // CONFIG_VP9_HIGHBITDEPTH
4293 if (x->source_variance < UINT_MAX) {
4294 const int var_adjust = (x->source_variance < 16);
4295 scale -= var_adjust;
4297 if (ref_frame > INTRA_FRAME && distortion2 * scale < qstep * qstep) {
4304 /* keep record of best compound/single-only prediction */
4305 if (!disable_skip && ref_frame != INTRA_FRAME) {
4306 int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
4308 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
4309 single_rate = rate2 - compmode_cost;
4310 hybrid_rate = rate2;
4312 single_rate = rate2;
4313 hybrid_rate = rate2 + compmode_cost;
4316 single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
4317 hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
4319 if (!comp_pred && single_rd < best_pred_rd[SINGLE_REFERENCE])
4320 best_pred_rd[SINGLE_REFERENCE] = single_rd;
4321 else if (comp_pred && single_rd < best_pred_rd[COMPOUND_REFERENCE])
4322 best_pred_rd[COMPOUND_REFERENCE] = single_rd;
4324 if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
4325 best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
4328 /* keep record of best filter type */
4329 if (!mode_excluded && !disable_skip && ref_frame != INTRA_FRAME &&
4330 cm->interp_filter != BILINEAR) {
4332 filter_cache[cm->interp_filter == SWITCHABLE ? SWITCHABLE_FILTERS
4333 : cm->interp_filter];
4335 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4336 if (ref == INT64_MAX)
4338 else if (filter_cache[i] == INT64_MAX)
4339 // when early termination is triggered, the encoder does not have
4340 // access to the rate-distortion cost. it only knows that the cost
4341 // should be above the maximum valid value. hence it takes the known
4342 // maximum plus an arbitrary constant as the rate-distortion cost.
4343 adj_rd = mask_filter - ref + 10;
4345 adj_rd = filter_cache[i] - ref;
4348 best_filter_rd[i] = VPXMIN(best_filter_rd[i], adj_rd);
4352 if (early_term) break;
4354 if (x->skip && !comp_pred) break;
4357 if (best_rd >= best_rd_so_far) {
4358 rd_cost->rate = INT_MAX;
4359 rd_cost->rdcost = INT64_MAX;
4363 // If we used an estimate for the uv intra rd in the loop above...
4364 if (sf->use_uv_intra_rd_estimate) {
4365 // Do Intra UV best rd mode selection if best mode choice above was intra.
4366 if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
4368 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra, &rate_uv_tokenonly,
4369 &dist_uv, &skip_uv, BLOCK_8X8, TX_4X4);
4373 if (best_rd == INT64_MAX) {
4374 rd_cost->rate = INT_MAX;
4375 rd_cost->dist = INT64_MAX;
4376 rd_cost->rdcost = INT64_MAX;
4380 assert((cm->interp_filter == SWITCHABLE) ||
4381 (cm->interp_filter == best_mbmode.interp_filter) ||
4382 !is_inter_block(&best_mbmode));
4384 vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact, sf->adaptive_rd_thresh,
4385 bsize, best_ref_index);
4389 x->skip |= best_skip2;
4390 if (!is_inter_block(&best_mbmode)) {
4391 for (i = 0; i < 4; i++) xd->mi[0]->bmi[i].as_mode = best_bmodes[i].as_mode;
4393 for (i = 0; i < 4; ++i)
4394 memcpy(&xd->mi[0]->bmi[i], &best_bmodes[i], sizeof(b_mode_info));
4396 mi->mv[0].as_int = xd->mi[0]->bmi[3].as_mv[0].as_int;
4397 mi->mv[1].as_int = xd->mi[0]->bmi[3].as_mv[1].as_int;
4400 for (i = 0; i < REFERENCE_MODES; ++i) {
4401 if (best_pred_rd[i] == INT64_MAX)
4402 best_pred_diff[i] = INT_MIN;
4404 best_pred_diff[i] = best_rd - best_pred_rd[i];
4408 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4409 if (best_filter_rd[i] == INT64_MAX)
4410 best_filter_diff[i] = 0;
4412 best_filter_diff[i] = best_rd - best_filter_rd[i];
4414 if (cm->interp_filter == SWITCHABLE)
4415 assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
4417 vp9_zero(best_filter_diff);
4420 store_coding_context(x, ctx, best_ref_index, best_pred_diff, best_filter_diff,