1 /*
2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11 #include <assert.h>
12 #include <math.h>
13
14 #include "./vp9_rtcd.h"
15 #include "./vpx_dsp_rtcd.h"
16
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"
21
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"
33
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"
44
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))
51
52 #define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | 0x01)
53
54 #define MIN_EARLY_TERM_INDEX 3
55 #define NEW_MV_DISCOUNT_FACTOR 8
56
57 typedef struct {
58 PREDICTION_MODE mode;
59 MV_REFERENCE_FRAME ref_frame[2];
60 } MODE_DEFINITION;
61
62 typedef struct { MV_REFERENCE_FRAME ref_frame[2]; } REF_DEFINITION;
63
64 struct rdcost_block_args {
65 const VP9_COMP *cpi;
66 MACROBLOCK *x;
67 ENTROPY_CONTEXT t_above[16];
68 ENTROPY_CONTEXT t_left[16];
69 int this_rate;
70 int64_t this_dist;
71 int64_t this_sse;
72 int64_t this_rd;
73 int64_t best_rd;
74 int exit_early;
75 int use_fast_coef_costing;
76 const scan_order *so;
77 uint8_t skippable;
78 };
79
80 #define LAST_NEW_MV_INDEX 6
81 static const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
82 { NEARESTMV, { LAST_FRAME, NONE } },
83 { NEARESTMV, { ALTREF_FRAME, NONE } },
84 { NEARESTMV, { GOLDEN_FRAME, NONE } },
85
86 { DC_PRED, { INTRA_FRAME, NONE } },
87
88 { NEWMV, { LAST_FRAME, NONE } },
89 { NEWMV, { ALTREF_FRAME, NONE } },
90 { NEWMV, { GOLDEN_FRAME, NONE } },
91
92 { NEARMV, { LAST_FRAME, NONE } },
93 { NEARMV, { ALTREF_FRAME, NONE } },
94 { NEARMV, { GOLDEN_FRAME, NONE } },
95
96 { ZEROMV, { LAST_FRAME, NONE } },
97 { ZEROMV, { GOLDEN_FRAME, NONE } },
98 { ZEROMV, { ALTREF_FRAME, NONE } },
99
100 { NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
101 { NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
102
103 { TM_PRED, { INTRA_FRAME, NONE } },
104
105 { NEARMV, { LAST_FRAME, ALTREF_FRAME } },
106 { NEWMV, { LAST_FRAME, ALTREF_FRAME } },
107 { NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
108 { NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
109
110 { ZEROMV, { LAST_FRAME, ALTREF_FRAME } },
111 { ZEROMV, { GOLDEN_FRAME, ALTREF_FRAME } },
112
113 { H_PRED, { INTRA_FRAME, NONE } },
114 { V_PRED, { INTRA_FRAME, NONE } },
115 { D135_PRED, { INTRA_FRAME, NONE } },
116 { D207_PRED, { INTRA_FRAME, NONE } },
117 { D153_PRED, { INTRA_FRAME, NONE } },
118 { D63_PRED, { INTRA_FRAME, NONE } },
119 { D117_PRED, { INTRA_FRAME, NONE } },
120 { D45_PRED, { INTRA_FRAME, NONE } },
121 };
122
123 static const REF_DEFINITION vp9_ref_order[MAX_REFS] = {
124 { { LAST_FRAME, NONE } }, { { GOLDEN_FRAME, NONE } },
125 { { ALTREF_FRAME, NONE } }, { { LAST_FRAME, ALTREF_FRAME } },
126 { { GOLDEN_FRAME, ALTREF_FRAME } }, { { INTRA_FRAME, NONE } },
127 };
128
swap_block_ptr(MACROBLOCK * x,PICK_MODE_CONTEXT * ctx,int m,int n,int min_plane,int max_plane)129 static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx, int m, int n,
130 int min_plane, int max_plane) {
131 int i;
132
133 for (i = min_plane; i < max_plane; ++i) {
134 struct macroblock_plane *const p = &x->plane[i];
135 struct macroblockd_plane *const pd = &x->e_mbd.plane[i];
136
137 p->coeff = ctx->coeff_pbuf[i][m];
138 p->qcoeff = ctx->qcoeff_pbuf[i][m];
139 pd->dqcoeff = ctx->dqcoeff_pbuf[i][m];
140 p->eobs = ctx->eobs_pbuf[i][m];
141
142 ctx->coeff_pbuf[i][m] = ctx->coeff_pbuf[i][n];
143 ctx->qcoeff_pbuf[i][m] = ctx->qcoeff_pbuf[i][n];
144 ctx->dqcoeff_pbuf[i][m] = ctx->dqcoeff_pbuf[i][n];
145 ctx->eobs_pbuf[i][m] = ctx->eobs_pbuf[i][n];
146
147 ctx->coeff_pbuf[i][n] = p->coeff;
148 ctx->qcoeff_pbuf[i][n] = p->qcoeff;
149 ctx->dqcoeff_pbuf[i][n] = pd->dqcoeff;
150 ctx->eobs_pbuf[i][n] = p->eobs;
151 }
152 }
153
model_rd_for_sb(VP9_COMP * cpi,BLOCK_SIZE bsize,MACROBLOCK * x,MACROBLOCKD * xd,int * out_rate_sum,int64_t * out_dist_sum,int * skip_txfm_sb,int64_t * skip_sse_sb)154 static void model_rd_for_sb(VP9_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x,
155 MACROBLOCKD *xd, int *out_rate_sum,
156 int64_t *out_dist_sum, int *skip_txfm_sb,
157 int64_t *skip_sse_sb) {
158 // Note our transform coeffs are 8 times an orthogonal transform.
159 // Hence quantizer step is also 8 times. To get effective quantizer
160 // we need to divide by 8 before sending to modeling function.
161 int i;
162 int64_t rate_sum = 0;
163 int64_t dist_sum = 0;
164 const int ref = xd->mi[0]->ref_frame[0];
165 unsigned int sse;
166 unsigned int var = 0;
167 unsigned int sum_sse = 0;
168 int64_t total_sse = 0;
169 int skip_flag = 1;
170 const int shift = 6;
171 int rate;
172 int64_t dist;
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
177 3;
178
179 x->pred_sse[ref] = 0;
180
181 for (i = 0; i < MAX_MB_PLANE; ++i) {
182 struct macroblock_plane *const p = &x->plane[i];
183 struct macroblockd_plane *const pd = &xd->plane[i];
184 const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
185 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
186 const BLOCK_SIZE unit_size = txsize_to_bsize[max_tx_size];
187 const int64_t dc_thr = p->quant_thred[0] >> shift;
188 const int64_t ac_thr = p->quant_thred[1] >> shift;
189 // The low thresholds are used to measure if the prediction errors are
190 // low enough so that we can skip the mode search.
191 const int64_t low_dc_thr = VPXMIN(50, dc_thr >> 2);
192 const int64_t low_ac_thr = VPXMIN(80, ac_thr >> 2);
193 int bw = 1 << (b_width_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
194 int bh = 1 << (b_height_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
195 int idx, idy;
196 int lw = b_width_log2_lookup[unit_size] + 2;
197 int lh = b_height_log2_lookup[unit_size] + 2;
198
199 sum_sse = 0;
200
201 for (idy = 0; idy < bh; ++idy) {
202 for (idx = 0; idx < bw; ++idx) {
203 uint8_t *src = p->src.buf + (idy * p->src.stride << lh) + (idx << lw);
204 uint8_t *dst = pd->dst.buf + (idy * pd->dst.stride << lh) + (idx << lh);
205 int block_idx = (idy << 1) + idx;
206 int low_err_skip = 0;
207
208 var = cpi->fn_ptr[unit_size].vf(src, p->src.stride, dst, pd->dst.stride,
209 &sse);
210 x->bsse[(i << 2) + block_idx] = sse;
211 sum_sse += sse;
212
213 x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_NONE;
214 if (!x->select_tx_size) {
215 // Check if all ac coefficients can be quantized to zero.
216 if (var < ac_thr || var == 0) {
217 x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_ONLY;
218
219 // Check if dc coefficient can be quantized to zero.
220 if (sse - var < dc_thr || sse == var) {
221 x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_DC;
222
223 if (!sse || (var < low_ac_thr && sse - var < low_dc_thr))
224 low_err_skip = 1;
225 }
226 }
227 }
228
229 if (skip_flag && !low_err_skip) skip_flag = 0;
230
231 if (i == 0) x->pred_sse[ref] += sse;
232 }
233 }
234
235 total_sse += sum_sse;
236
237 // Fast approximate the modelling function.
238 if (cpi->sf.simple_model_rd_from_var) {
239 int64_t rate;
240 const int64_t square_error = sum_sse;
241 int quantizer = (pd->dequant[1] >> dequant_shift);
242
243 if (quantizer < 120)
244 rate = (square_error * (280 - quantizer)) >> (16 - VP9_PROB_COST_SHIFT);
245 else
246 rate = 0;
247 dist = (square_error * quantizer) >> 8;
248 rate_sum += rate;
249 dist_sum += dist;
250 } else {
251 vp9_model_rd_from_var_lapndz(sum_sse, num_pels_log2_lookup[bs],
252 pd->dequant[1] >> dequant_shift, &rate,
253 &dist);
254 rate_sum += rate;
255 dist_sum += dist;
256 }
257 }
258
259 *skip_txfm_sb = skip_flag;
260 *skip_sse_sb = total_sse << 4;
261 *out_rate_sum = (int)rate_sum;
262 *out_dist_sum = dist_sum << 4;
263 }
264
265 #if CONFIG_VP9_HIGHBITDEPTH
vp9_highbd_block_error_c(const tran_low_t * coeff,const tran_low_t * dqcoeff,intptr_t block_size,int64_t * ssz,int bd)266 int64_t vp9_highbd_block_error_c(const tran_low_t *coeff,
267 const tran_low_t *dqcoeff, intptr_t block_size,
268 int64_t *ssz, int bd) {
269 int i;
270 int64_t error = 0, sqcoeff = 0;
271 int shift = 2 * (bd - 8);
272 int rounding = shift > 0 ? 1 << (shift - 1) : 0;
273
274 for (i = 0; i < block_size; i++) {
275 const int64_t diff = coeff[i] - dqcoeff[i];
276 error += diff * diff;
277 sqcoeff += (int64_t)coeff[i] * (int64_t)coeff[i];
278 }
279 assert(error >= 0 && sqcoeff >= 0);
280 error = (error + rounding) >> shift;
281 sqcoeff = (sqcoeff + rounding) >> shift;
282
283 *ssz = sqcoeff;
284 return error;
285 }
286
vp9_highbd_block_error_8bit_c(const tran_low_t * coeff,const tran_low_t * dqcoeff,intptr_t block_size,int64_t * ssz)287 int64_t vp9_highbd_block_error_8bit_c(const tran_low_t *coeff,
288 const tran_low_t *dqcoeff,
289 intptr_t block_size, int64_t *ssz) {
290 // Note that the C versions of these 2 functions (vp9_block_error and
291 // vp9_highbd_block_error_8bit are the same, but the optimized assembly
292 // routines are not compatible in the non high bitdepth configuration, so
293 // they still cannot share the same name.
294 return vp9_block_error_c(coeff, dqcoeff, block_size, ssz);
295 }
296
vp9_highbd_block_error_dispatch(const tran_low_t * coeff,const tran_low_t * dqcoeff,intptr_t block_size,int64_t * ssz,int bd)297 static int64_t vp9_highbd_block_error_dispatch(const tran_low_t *coeff,
298 const tran_low_t *dqcoeff,
299 intptr_t block_size,
300 int64_t *ssz, int bd) {
301 if (bd == 8) {
302 return vp9_highbd_block_error_8bit(coeff, dqcoeff, block_size, ssz);
303 } else {
304 return vp9_highbd_block_error(coeff, dqcoeff, block_size, ssz, bd);
305 }
306 }
307 #endif // CONFIG_VP9_HIGHBITDEPTH
308
vp9_block_error_c(const tran_low_t * coeff,const tran_low_t * dqcoeff,intptr_t block_size,int64_t * ssz)309 int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
310 intptr_t block_size, int64_t *ssz) {
311 int i;
312 int64_t error = 0, sqcoeff = 0;
313
314 for (i = 0; i < block_size; i++) {
315 const int diff = coeff[i] - dqcoeff[i];
316 error += diff * diff;
317 sqcoeff += coeff[i] * coeff[i];
318 }
319
320 *ssz = sqcoeff;
321 return error;
322 }
323
vp9_block_error_fp_c(const int16_t * coeff,const int16_t * dqcoeff,int block_size)324 int64_t vp9_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff,
325 int block_size) {
326 int i;
327 int64_t error = 0;
328
329 for (i = 0; i < block_size; i++) {
330 const int diff = coeff[i] - dqcoeff[i];
331 error += diff * diff;
332 }
333
334 return error;
335 }
336
337 /* The trailing '0' is a terminator which is used inside cost_coeffs() to
338 * decide whether to include cost of a trailing EOB node or not (i.e. we
339 * can skip this if the last coefficient in this transform block, e.g. the
340 * 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block,
341 * were non-zero). */
342 static const int16_t band_counts[TX_SIZES][8] = {
343 { 1, 2, 3, 4, 3, 16 - 13, 0 },
344 { 1, 2, 3, 4, 11, 64 - 21, 0 },
345 { 1, 2, 3, 4, 11, 256 - 21, 0 },
346 { 1, 2, 3, 4, 11, 1024 - 21, 0 },
347 };
cost_coeffs(MACROBLOCK * x,int plane,int block,TX_SIZE tx_size,int pt,const int16_t * scan,const int16_t * nb,int use_fast_coef_costing)348 static int cost_coeffs(MACROBLOCK *x, int plane, int block, TX_SIZE tx_size,
349 int pt, const int16_t *scan, const int16_t *nb,
350 int use_fast_coef_costing) {
351 MACROBLOCKD *const xd = &x->e_mbd;
352 MODE_INFO *mi = xd->mi[0];
353 const struct macroblock_plane *p = &x->plane[plane];
354 const PLANE_TYPE type = get_plane_type(plane);
355 const int16_t *band_count = &band_counts[tx_size][1];
356 const int eob = p->eobs[block];
357 const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
358 unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
359 x->token_costs[tx_size][type][is_inter_block(mi)];
360 uint8_t token_cache[32 * 32];
361 int c, cost;
362 #if CONFIG_VP9_HIGHBITDEPTH
363 const int *cat6_high_cost = vp9_get_high_cost_table(xd->bd);
364 #else
365 const int *cat6_high_cost = vp9_get_high_cost_table(8);
366 #endif
367
368 // Check for consistency of tx_size with mode info
369 assert(type == PLANE_TYPE_Y
370 ? mi->tx_size == tx_size
371 : get_uv_tx_size(mi, &xd->plane[plane]) == tx_size);
372
373 if (eob == 0) {
374 // single eob token
375 cost = token_costs[0][0][pt][EOB_TOKEN];
376 c = 0;
377 } else {
378 if (use_fast_coef_costing) {
379 int band_left = *band_count++;
380
381 // dc token
382 int v = qcoeff[0];
383 int16_t prev_t;
384 cost = vp9_get_token_cost(v, &prev_t, cat6_high_cost);
385 cost += (*token_costs)[0][pt][prev_t];
386
387 token_cache[0] = vp9_pt_energy_class[prev_t];
388 ++token_costs;
389
390 // ac tokens
391 for (c = 1; c < eob; c++) {
392 const int rc = scan[c];
393 int16_t t;
394
395 v = qcoeff[rc];
396 cost += vp9_get_token_cost(v, &t, cat6_high_cost);
397 cost += (*token_costs)[!prev_t][!prev_t][t];
398 prev_t = t;
399 if (!--band_left) {
400 band_left = *band_count++;
401 ++token_costs;
402 }
403 }
404
405 // eob token
406 if (band_left) cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
407
408 } else { // !use_fast_coef_costing
409 int band_left = *band_count++;
410
411 // dc token
412 int v = qcoeff[0];
413 int16_t tok;
414 unsigned int(*tok_cost_ptr)[COEFF_CONTEXTS][ENTROPY_TOKENS];
415 cost = vp9_get_token_cost(v, &tok, cat6_high_cost);
416 cost += (*token_costs)[0][pt][tok];
417
418 token_cache[0] = vp9_pt_energy_class[tok];
419 ++token_costs;
420
421 tok_cost_ptr = &((*token_costs)[!tok]);
422
423 // ac tokens
424 for (c = 1; c < eob; c++) {
425 const int rc = scan[c];
426
427 v = qcoeff[rc];
428 cost += vp9_get_token_cost(v, &tok, cat6_high_cost);
429 pt = get_coef_context(nb, token_cache, c);
430 cost += (*tok_cost_ptr)[pt][tok];
431 token_cache[rc] = vp9_pt_energy_class[tok];
432 if (!--band_left) {
433 band_left = *band_count++;
434 ++token_costs;
435 }
436 tok_cost_ptr = &((*token_costs)[!tok]);
437 }
438
439 // eob token
440 if (band_left) {
441 pt = get_coef_context(nb, token_cache, c);
442 cost += (*token_costs)[0][pt][EOB_TOKEN];
443 }
444 }
445 }
446
447 return cost;
448 }
449
num_4x4_to_edge(int plane_4x4_dim,int mb_to_edge_dim,int subsampling_dim,int blk_dim)450 static INLINE int num_4x4_to_edge(int plane_4x4_dim, int mb_to_edge_dim,
451 int subsampling_dim, int blk_dim) {
452 return plane_4x4_dim + (mb_to_edge_dim >> (5 + subsampling_dim)) - blk_dim;
453 }
454
455 // Compute the pixel domain sum square error on all visible 4x4s in the
456 // transform block.
pixel_sse(const VP9_COMP * const cpi,const MACROBLOCKD * xd,const struct macroblockd_plane * const pd,const uint8_t * src,const int src_stride,const uint8_t * dst,const int dst_stride,int blk_row,int blk_col,const BLOCK_SIZE plane_bsize,const BLOCK_SIZE tx_bsize)457 static unsigned pixel_sse(const VP9_COMP *const cpi, const MACROBLOCKD *xd,
458 const struct macroblockd_plane *const pd,
459 const uint8_t *src, const int src_stride,
460 const uint8_t *dst, const int dst_stride, int blk_row,
461 int blk_col, const BLOCK_SIZE plane_bsize,
462 const BLOCK_SIZE tx_bsize) {
463 unsigned int sse = 0;
464 const int plane_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
465 const int plane_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
466 const int tx_4x4_w = num_4x4_blocks_wide_lookup[tx_bsize];
467 const int tx_4x4_h = num_4x4_blocks_high_lookup[tx_bsize];
468 int b4x4s_to_right_edge = num_4x4_to_edge(plane_4x4_w, xd->mb_to_right_edge,
469 pd->subsampling_x, blk_col);
470 int b4x4s_to_bottom_edge = num_4x4_to_edge(plane_4x4_h, xd->mb_to_bottom_edge,
471 pd->subsampling_y, blk_row);
472 if (tx_bsize == BLOCK_4X4 ||
473 (b4x4s_to_right_edge >= tx_4x4_w && b4x4s_to_bottom_edge >= tx_4x4_h)) {
474 cpi->fn_ptr[tx_bsize].vf(src, src_stride, dst, dst_stride, &sse);
475 } else {
476 const vpx_variance_fn_t vf_4x4 = cpi->fn_ptr[BLOCK_4X4].vf;
477 int r, c;
478 unsigned this_sse = 0;
479 int max_r = VPXMIN(b4x4s_to_bottom_edge, tx_4x4_h);
480 int max_c = VPXMIN(b4x4s_to_right_edge, tx_4x4_w);
481 sse = 0;
482 // if we are in the unrestricted motion border.
483 for (r = 0; r < max_r; ++r) {
484 // Skip visiting the sub blocks that are wholly within the UMV.
485 for (c = 0; c < max_c; ++c) {
486 vf_4x4(src + r * src_stride * 4 + c * 4, src_stride,
487 dst + r * dst_stride * 4 + c * 4, dst_stride, &this_sse);
488 sse += this_sse;
489 }
490 }
491 }
492 return sse;
493 }
494
495 // Compute the squares sum squares on all visible 4x4s in the transform block.
sum_squares_visible(const MACROBLOCKD * xd,const struct macroblockd_plane * const pd,const int16_t * diff,const int diff_stride,int blk_row,int blk_col,const BLOCK_SIZE plane_bsize,const BLOCK_SIZE tx_bsize)496 static int64_t sum_squares_visible(const MACROBLOCKD *xd,
497 const struct macroblockd_plane *const pd,
498 const int16_t *diff, const int diff_stride,
499 int blk_row, int blk_col,
500 const BLOCK_SIZE plane_bsize,
501 const BLOCK_SIZE tx_bsize) {
502 int64_t sse;
503 const int plane_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
504 const int plane_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
505 const int tx_4x4_w = num_4x4_blocks_wide_lookup[tx_bsize];
506 const int tx_4x4_h = num_4x4_blocks_high_lookup[tx_bsize];
507 int b4x4s_to_right_edge = num_4x4_to_edge(plane_4x4_w, xd->mb_to_right_edge,
508 pd->subsampling_x, blk_col);
509 int b4x4s_to_bottom_edge = num_4x4_to_edge(plane_4x4_h, xd->mb_to_bottom_edge,
510 pd->subsampling_y, blk_row);
511 if (tx_bsize == BLOCK_4X4 ||
512 (b4x4s_to_right_edge >= tx_4x4_w && b4x4s_to_bottom_edge >= tx_4x4_h)) {
513 sse = (int64_t)vpx_sum_squares_2d_i16(diff, diff_stride, tx_bsize);
514 } else {
515 int r, c;
516 int max_r = VPXMIN(b4x4s_to_bottom_edge, tx_4x4_h);
517 int max_c = VPXMIN(b4x4s_to_right_edge, tx_4x4_w);
518 sse = 0;
519 // if we are in the unrestricted motion border.
520 for (r = 0; r < max_r; ++r) {
521 // Skip visiting the sub blocks that are wholly within the UMV.
522 for (c = 0; c < max_c; ++c) {
523 sse += (int64_t)vpx_sum_squares_2d_i16(diff, diff_stride, BLOCK_4X4);
524 }
525 }
526 }
527 return sse;
528 }
529
dist_block(const VP9_COMP * cpi,MACROBLOCK * x,int plane,BLOCK_SIZE plane_bsize,int block,int blk_row,int blk_col,TX_SIZE tx_size,int64_t * out_dist,int64_t * out_sse)530 static void dist_block(const VP9_COMP *cpi, MACROBLOCK *x, int plane,
531 BLOCK_SIZE plane_bsize, int block, int blk_row,
532 int blk_col, TX_SIZE tx_size, int64_t *out_dist,
533 int64_t *out_sse) {
534 MACROBLOCKD *const xd = &x->e_mbd;
535 const struct macroblock_plane *const p = &x->plane[plane];
536 const struct macroblockd_plane *const pd = &xd->plane[plane];
537
538 if (x->block_tx_domain) {
539 const int ss_txfrm_size = tx_size << 1;
540 int64_t this_sse;
541 const int shift = tx_size == TX_32X32 ? 0 : 2;
542 const tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
543 const tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
544 #if CONFIG_VP9_HIGHBITDEPTH
545 const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
546 *out_dist = vp9_highbd_block_error_dispatch(
547 coeff, dqcoeff, 16 << ss_txfrm_size, &this_sse, bd) >>
548 shift;
549 #else
550 *out_dist =
551 vp9_block_error(coeff, dqcoeff, 16 << ss_txfrm_size, &this_sse) >>
552 shift;
553 #endif // CONFIG_VP9_HIGHBITDEPTH
554 *out_sse = this_sse >> shift;
555
556 if (x->skip_encode && !is_inter_block(xd->mi[0])) {
557 // TODO(jingning): tune the model to better capture the distortion.
558 const int64_t p =
559 (pd->dequant[1] * pd->dequant[1] * (1 << ss_txfrm_size)) >>
560 #if CONFIG_VP9_HIGHBITDEPTH
561 (shift + 2 + (bd - 8) * 2);
562 #else
563 (shift + 2);
564 #endif // CONFIG_VP9_HIGHBITDEPTH
565 *out_dist += (p >> 4);
566 *out_sse += p;
567 }
568 } else {
569 const BLOCK_SIZE tx_bsize = txsize_to_bsize[tx_size];
570 const int bs = 4 * num_4x4_blocks_wide_lookup[tx_bsize];
571 const int src_stride = p->src.stride;
572 const int dst_stride = pd->dst.stride;
573 const int src_idx = 4 * (blk_row * src_stride + blk_col);
574 const int dst_idx = 4 * (blk_row * dst_stride + blk_col);
575 const uint8_t *src = &p->src.buf[src_idx];
576 const uint8_t *dst = &pd->dst.buf[dst_idx];
577 const tran_low_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
578 const uint16_t *eob = &p->eobs[block];
579 unsigned int tmp;
580
581 tmp = pixel_sse(cpi, xd, pd, src, src_stride, dst, dst_stride, blk_row,
582 blk_col, plane_bsize, tx_bsize);
583 *out_sse = (int64_t)tmp * 16;
584
585 if (*eob) {
586 #if CONFIG_VP9_HIGHBITDEPTH
587 DECLARE_ALIGNED(16, uint16_t, recon16[1024]);
588 uint8_t *recon = (uint8_t *)recon16;
589 #else
590 DECLARE_ALIGNED(16, uint8_t, recon[1024]);
591 #endif // CONFIG_VP9_HIGHBITDEPTH
592
593 #if CONFIG_VP9_HIGHBITDEPTH
594 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
595 recon = CONVERT_TO_BYTEPTR(recon);
596 vpx_highbd_convolve_copy(dst, dst_stride, recon, 32, NULL, 0, NULL, 0,
597 bs, bs, xd->bd);
598 if (xd->lossless) {
599 vp9_highbd_iwht4x4_add(dqcoeff, recon, 32, *eob, xd->bd);
600 } else {
601 switch (tx_size) {
602 case TX_4X4:
603 vp9_highbd_idct4x4_add(dqcoeff, recon, 32, *eob, xd->bd);
604 break;
605 case TX_8X8:
606 vp9_highbd_idct8x8_add(dqcoeff, recon, 32, *eob, xd->bd);
607 break;
608 case TX_16X16:
609 vp9_highbd_idct16x16_add(dqcoeff, recon, 32, *eob, xd->bd);
610 break;
611 case TX_32X32:
612 vp9_highbd_idct32x32_add(dqcoeff, recon, 32, *eob, xd->bd);
613 break;
614 default: assert(0 && "Invalid transform size");
615 }
616 }
617 } else {
618 #endif // CONFIG_VP9_HIGHBITDEPTH
619 vpx_convolve_copy(dst, dst_stride, recon, 32, NULL, 0, NULL, 0, bs, bs);
620 switch (tx_size) {
621 case TX_32X32: vp9_idct32x32_add(dqcoeff, recon, 32, *eob); break;
622 case TX_16X16: vp9_idct16x16_add(dqcoeff, recon, 32, *eob); break;
623 case TX_8X8: vp9_idct8x8_add(dqcoeff, recon, 32, *eob); break;
624 case TX_4X4:
625 // this is like vp9_short_idct4x4 but has a special case around
626 // eob<=1, which is significant (not just an optimization) for
627 // the lossless case.
628 x->itxm_add(dqcoeff, recon, 32, *eob);
629 break;
630 default: assert(0 && "Invalid transform size"); break;
631 }
632 #if CONFIG_VP9_HIGHBITDEPTH
633 }
634 #endif // CONFIG_VP9_HIGHBITDEPTH
635
636 tmp = pixel_sse(cpi, xd, pd, src, src_stride, recon, 32, blk_row, blk_col,
637 plane_bsize, tx_bsize);
638 }
639
640 *out_dist = (int64_t)tmp * 16;
641 }
642 }
643
rate_block(int plane,int block,TX_SIZE tx_size,int coeff_ctx,struct rdcost_block_args * args)644 static int rate_block(int plane, int block, TX_SIZE tx_size, int coeff_ctx,
645 struct rdcost_block_args *args) {
646 return cost_coeffs(args->x, plane, block, tx_size, coeff_ctx, args->so->scan,
647 args->so->neighbors, args->use_fast_coef_costing);
648 }
649
block_rd_txfm(int plane,int block,int blk_row,int blk_col,BLOCK_SIZE plane_bsize,TX_SIZE tx_size,void * arg)650 static void block_rd_txfm(int plane, int block, int blk_row, int blk_col,
651 BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
652 struct rdcost_block_args *args = arg;
653 MACROBLOCK *const x = args->x;
654 MACROBLOCKD *const xd = &x->e_mbd;
655 MODE_INFO *const mi = xd->mi[0];
656 int64_t rd1, rd2, rd;
657 int rate;
658 int64_t dist;
659 int64_t sse;
660 const int coeff_ctx =
661 combine_entropy_contexts(args->t_left[blk_row], args->t_above[blk_col]);
662
663 if (args->exit_early) return;
664
665 if (!is_inter_block(mi)) {
666 struct encode_b_args intra_arg = { x, x->block_qcoeff_opt, args->t_above,
667 args->t_left, &mi->skip };
668 vp9_encode_block_intra(plane, block, blk_row, blk_col, plane_bsize, tx_size,
669 &intra_arg);
670 if (x->block_tx_domain) {
671 dist_block(args->cpi, x, plane, plane_bsize, block, blk_row, blk_col,
672 tx_size, &dist, &sse);
673 } else {
674 const BLOCK_SIZE tx_bsize = txsize_to_bsize[tx_size];
675 const struct macroblock_plane *const p = &x->plane[plane];
676 const struct macroblockd_plane *const pd = &xd->plane[plane];
677 const int src_stride = p->src.stride;
678 const int dst_stride = pd->dst.stride;
679 const int diff_stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
680 const uint8_t *src = &p->src.buf[4 * (blk_row * src_stride + blk_col)];
681 const uint8_t *dst = &pd->dst.buf[4 * (blk_row * dst_stride + blk_col)];
682 const int16_t *diff = &p->src_diff[4 * (blk_row * diff_stride + blk_col)];
683 unsigned int tmp;
684 sse = sum_squares_visible(xd, pd, diff, diff_stride, blk_row, blk_col,
685 plane_bsize, tx_bsize);
686 #if CONFIG_VP9_HIGHBITDEPTH
687 if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && (xd->bd > 8))
688 sse = ROUND64_POWER_OF_TWO(sse, (xd->bd - 8) * 2);
689 #endif // CONFIG_VP9_HIGHBITDEPTH
690 sse = sse * 16;
691 tmp = pixel_sse(args->cpi, xd, pd, src, src_stride, dst, dst_stride,
692 blk_row, blk_col, plane_bsize, tx_bsize);
693 dist = (int64_t)tmp * 16;
694 }
695 } else if (max_txsize_lookup[plane_bsize] == tx_size) {
696 if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] ==
697 SKIP_TXFM_NONE) {
698 // full forward transform and quantization
699 vp9_xform_quant(x, plane, block, blk_row, blk_col, plane_bsize, tx_size);
700 if (x->block_qcoeff_opt)
701 vp9_optimize_b(x, plane, block, tx_size, coeff_ctx);
702 dist_block(args->cpi, x, plane, plane_bsize, block, blk_row, blk_col,
703 tx_size, &dist, &sse);
704 } else if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] ==
705 SKIP_TXFM_AC_ONLY) {
706 // compute DC coefficient
707 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[plane].coeff, block);
708 tran_low_t *const dqcoeff = BLOCK_OFFSET(xd->plane[plane].dqcoeff, block);
709 vp9_xform_quant_dc(x, plane, block, blk_row, blk_col, plane_bsize,
710 tx_size);
711 sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
712 dist = sse;
713 if (x->plane[plane].eobs[block]) {
714 const int64_t orig_sse = (int64_t)coeff[0] * coeff[0];
715 const int64_t resd_sse = coeff[0] - dqcoeff[0];
716 int64_t dc_correct = orig_sse - resd_sse * resd_sse;
717 #if CONFIG_VP9_HIGHBITDEPTH
718 dc_correct >>= ((xd->bd - 8) * 2);
719 #endif
720 if (tx_size != TX_32X32) dc_correct >>= 2;
721
722 dist = VPXMAX(0, sse - dc_correct);
723 }
724 } else {
725 // SKIP_TXFM_AC_DC
726 // skip forward transform
727 x->plane[plane].eobs[block] = 0;
728 sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
729 dist = sse;
730 }
731 } else {
732 // full forward transform and quantization
733 vp9_xform_quant(x, plane, block, blk_row, blk_col, plane_bsize, tx_size);
734 if (x->block_qcoeff_opt)
735 vp9_optimize_b(x, plane, block, tx_size, coeff_ctx);
736 dist_block(args->cpi, x, plane, plane_bsize, block, blk_row, blk_col,
737 tx_size, &dist, &sse);
738 }
739
740 rd = RDCOST(x->rdmult, x->rddiv, 0, dist);
741 if (args->this_rd + rd > args->best_rd) {
742 args->exit_early = 1;
743 return;
744 }
745
746 rate = rate_block(plane, block, tx_size, coeff_ctx, args);
747 args->t_above[blk_col] = (x->plane[plane].eobs[block] > 0) ? 1 : 0;
748 args->t_left[blk_row] = (x->plane[plane].eobs[block] > 0) ? 1 : 0;
749 rd1 = RDCOST(x->rdmult, x->rddiv, rate, dist);
750 rd2 = RDCOST(x->rdmult, x->rddiv, 0, sse);
751
752 // TODO(jingning): temporarily enabled only for luma component
753 rd = VPXMIN(rd1, rd2);
754 if (plane == 0)
755 x->zcoeff_blk[tx_size][block] =
756 !x->plane[plane].eobs[block] || (rd1 > rd2 && !xd->lossless);
757
758 args->this_rate += rate;
759 args->this_dist += dist;
760 args->this_sse += sse;
761 args->this_rd += rd;
762
763 if (args->this_rd > args->best_rd) {
764 args->exit_early = 1;
765 return;
766 }
767
768 args->skippable &= !x->plane[plane].eobs[block];
769 }
770
txfm_rd_in_plane(const VP9_COMP * cpi,MACROBLOCK * x,int * rate,int64_t * distortion,int * skippable,int64_t * sse,int64_t ref_best_rd,int plane,BLOCK_SIZE bsize,TX_SIZE tx_size,int use_fast_coef_casting)771 static void txfm_rd_in_plane(const VP9_COMP *cpi, MACROBLOCK *x, int *rate,
772 int64_t *distortion, int *skippable, int64_t *sse,
773 int64_t ref_best_rd, int plane, BLOCK_SIZE bsize,
774 TX_SIZE tx_size, int use_fast_coef_casting) {
775 MACROBLOCKD *const xd = &x->e_mbd;
776 const struct macroblockd_plane *const pd = &xd->plane[plane];
777 struct rdcost_block_args args;
778 vp9_zero(args);
779 args.cpi = cpi;
780 args.x = x;
781 args.best_rd = ref_best_rd;
782 args.use_fast_coef_costing = use_fast_coef_casting;
783 args.skippable = 1;
784
785 if (plane == 0) xd->mi[0]->tx_size = tx_size;
786
787 vp9_get_entropy_contexts(bsize, tx_size, pd, args.t_above, args.t_left);
788
789 args.so = get_scan(xd, tx_size, get_plane_type(plane), 0);
790
791 vp9_foreach_transformed_block_in_plane(xd, bsize, plane, block_rd_txfm,
792 &args);
793 if (args.exit_early) {
794 *rate = INT_MAX;
795 *distortion = INT64_MAX;
796 *sse = INT64_MAX;
797 *skippable = 0;
798 } else {
799 *distortion = args.this_dist;
800 *rate = args.this_rate;
801 *sse = args.this_sse;
802 *skippable = args.skippable;
803 }
804 }
805
choose_largest_tx_size(VP9_COMP * cpi,MACROBLOCK * x,int * rate,int64_t * distortion,int * skip,int64_t * sse,int64_t ref_best_rd,BLOCK_SIZE bs)806 static void choose_largest_tx_size(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
807 int64_t *distortion, int *skip, int64_t *sse,
808 int64_t ref_best_rd, BLOCK_SIZE bs) {
809 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
810 VP9_COMMON *const cm = &cpi->common;
811 const TX_SIZE largest_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
812 MACROBLOCKD *const xd = &x->e_mbd;
813 MODE_INFO *const mi = xd->mi[0];
814
815 mi->tx_size = VPXMIN(max_tx_size, largest_tx_size);
816
817 txfm_rd_in_plane(cpi, x, rate, distortion, skip, sse, ref_best_rd, 0, bs,
818 mi->tx_size, cpi->sf.use_fast_coef_costing);
819 }
820
choose_tx_size_from_rd(VP9_COMP * cpi,MACROBLOCK * x,int * rate,int64_t * distortion,int * skip,int64_t * psse,int64_t ref_best_rd,BLOCK_SIZE bs)821 static void choose_tx_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
822 int64_t *distortion, int *skip,
823 int64_t *psse, int64_t ref_best_rd,
824 BLOCK_SIZE bs) {
825 const TX_SIZE max_tx_size = max_txsize_lookup[bs];
826 VP9_COMMON *const cm = &cpi->common;
827 MACROBLOCKD *const xd = &x->e_mbd;
828 MODE_INFO *const mi = xd->mi[0];
829 vpx_prob skip_prob = vp9_get_skip_prob(cm, xd);
830 int r[TX_SIZES][2], s[TX_SIZES];
831 int64_t d[TX_SIZES], sse[TX_SIZES];
832 int64_t rd[TX_SIZES][2] = { { INT64_MAX, INT64_MAX },
833 { INT64_MAX, INT64_MAX },
834 { INT64_MAX, INT64_MAX },
835 { INT64_MAX, INT64_MAX } };
836 int n, m;
837 int s0, s1;
838 int64_t best_rd = INT64_MAX;
839 TX_SIZE best_tx = max_tx_size;
840 int start_tx, end_tx;
841
842 const vpx_prob *tx_probs = get_tx_probs2(max_tx_size, xd, &cm->fc->tx_probs);
843 assert(skip_prob > 0);
844 s0 = vp9_cost_bit(skip_prob, 0);
845 s1 = vp9_cost_bit(skip_prob, 1);
846
847 if (cm->tx_mode == TX_MODE_SELECT) {
848 start_tx = max_tx_size;
849 end_tx = 0;
850 } else {
851 TX_SIZE chosen_tx_size =
852 VPXMIN(max_tx_size, tx_mode_to_biggest_tx_size[cm->tx_mode]);
853 start_tx = chosen_tx_size;
854 end_tx = chosen_tx_size;
855 }
856
857 for (n = start_tx; n >= end_tx; n--) {
858 int r_tx_size = 0;
859 for (m = 0; m <= n - (n == (int)max_tx_size); m++) {
860 if (m == n)
861 r_tx_size += vp9_cost_zero(tx_probs[m]);
862 else
863 r_tx_size += vp9_cost_one(tx_probs[m]);
864 }
865 txfm_rd_in_plane(cpi, x, &r[n][0], &d[n], &s[n], &sse[n], ref_best_rd, 0,
866 bs, n, cpi->sf.use_fast_coef_costing);
867 r[n][1] = r[n][0];
868 if (r[n][0] < INT_MAX) {
869 r[n][1] += r_tx_size;
870 }
871 if (d[n] == INT64_MAX || r[n][0] == INT_MAX) {
872 rd[n][0] = rd[n][1] = INT64_MAX;
873 } else if (s[n]) {
874 if (is_inter_block(mi)) {
875 rd[n][0] = rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1, sse[n]);
876 r[n][1] -= r_tx_size;
877 } else {
878 rd[n][0] = RDCOST(x->rdmult, x->rddiv, s1, sse[n]);
879 rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1 + r_tx_size, sse[n]);
880 }
881 } else {
882 rd[n][0] = RDCOST(x->rdmult, x->rddiv, r[n][0] + s0, d[n]);
883 rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]);
884 }
885
886 if (is_inter_block(mi) && !xd->lossless && !s[n] && sse[n] != INT64_MAX) {
887 rd[n][0] = VPXMIN(rd[n][0], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
888 rd[n][1] = VPXMIN(rd[n][1], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
889 }
890
891 // Early termination in transform size search.
892 if (cpi->sf.tx_size_search_breakout &&
893 (rd[n][1] == INT64_MAX ||
894 (n < (int)max_tx_size && rd[n][1] > rd[n + 1][1]) || s[n] == 1))
895 break;
896
897 if (rd[n][1] < best_rd) {
898 best_tx = n;
899 best_rd = rd[n][1];
900 }
901 }
902 mi->tx_size = best_tx;
903
904 *distortion = d[mi->tx_size];
905 *rate = r[mi->tx_size][cm->tx_mode == TX_MODE_SELECT];
906 *skip = s[mi->tx_size];
907 *psse = sse[mi->tx_size];
908 }
909
super_block_yrd(VP9_COMP * cpi,MACROBLOCK * x,int * rate,int64_t * distortion,int * skip,int64_t * psse,BLOCK_SIZE bs,int64_t ref_best_rd)910 static void super_block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
911 int64_t *distortion, int *skip, int64_t *psse,
912 BLOCK_SIZE bs, int64_t ref_best_rd) {
913 MACROBLOCKD *xd = &x->e_mbd;
914 int64_t sse;
915 int64_t *ret_sse = psse ? psse : &sse;
916
917 assert(bs == xd->mi[0]->sb_type);
918
919 if (cpi->sf.tx_size_search_method == USE_LARGESTALL || xd->lossless) {
920 choose_largest_tx_size(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
921 bs);
922 } else {
923 choose_tx_size_from_rd(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
924 bs);
925 }
926 }
927
conditional_skipintra(PREDICTION_MODE mode,PREDICTION_MODE best_intra_mode)928 static int conditional_skipintra(PREDICTION_MODE mode,
929 PREDICTION_MODE best_intra_mode) {
930 if (mode == D117_PRED && best_intra_mode != V_PRED &&
931 best_intra_mode != D135_PRED)
932 return 1;
933 if (mode == D63_PRED && best_intra_mode != V_PRED &&
934 best_intra_mode != D45_PRED)
935 return 1;
936 if (mode == D207_PRED && best_intra_mode != H_PRED &&
937 best_intra_mode != D45_PRED)
938 return 1;
939 if (mode == D153_PRED && best_intra_mode != H_PRED &&
940 best_intra_mode != D135_PRED)
941 return 1;
942 return 0;
943 }
944
rd_pick_intra4x4block(VP9_COMP * cpi,MACROBLOCK * x,int row,int col,PREDICTION_MODE * best_mode,const int * bmode_costs,ENTROPY_CONTEXT * a,ENTROPY_CONTEXT * l,int * bestrate,int * bestratey,int64_t * bestdistortion,BLOCK_SIZE bsize,int64_t rd_thresh)945 static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x, int row,
946 int col, PREDICTION_MODE *best_mode,
947 const int *bmode_costs, ENTROPY_CONTEXT *a,
948 ENTROPY_CONTEXT *l, int *bestrate,
949 int *bestratey, int64_t *bestdistortion,
950 BLOCK_SIZE bsize, int64_t rd_thresh) {
951 PREDICTION_MODE mode;
952 MACROBLOCKD *const xd = &x->e_mbd;
953 int64_t best_rd = rd_thresh;
954 struct macroblock_plane *p = &x->plane[0];
955 struct macroblockd_plane *pd = &xd->plane[0];
956 const int src_stride = p->src.stride;
957 const int dst_stride = pd->dst.stride;
958 const uint8_t *src_init = &p->src.buf[row * 4 * src_stride + col * 4];
959 uint8_t *dst_init = &pd->dst.buf[row * 4 * src_stride + col * 4];
960 ENTROPY_CONTEXT ta[2], tempa[2];
961 ENTROPY_CONTEXT tl[2], templ[2];
962 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
963 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
964 int idx, idy;
965 uint8_t best_dst[8 * 8];
966 #if CONFIG_VP9_HIGHBITDEPTH
967 uint16_t best_dst16[8 * 8];
968 #endif
969 memcpy(ta, a, num_4x4_blocks_wide * sizeof(a[0]));
970 memcpy(tl, l, num_4x4_blocks_high * sizeof(l[0]));
971
972 xd->mi[0]->tx_size = TX_4X4;
973
974 #if CONFIG_VP9_HIGHBITDEPTH
975 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
976 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
977 int64_t this_rd;
978 int ratey = 0;
979 int64_t distortion = 0;
980 int rate = bmode_costs[mode];
981
982 if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode))) continue;
983
984 // Only do the oblique modes if the best so far is
985 // one of the neighboring directional modes
986 if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
987 if (conditional_skipintra(mode, *best_mode)) continue;
988 }
989
990 memcpy(tempa, ta, num_4x4_blocks_wide * sizeof(ta[0]));
991 memcpy(templ, tl, num_4x4_blocks_high * sizeof(tl[0]));
992
993 for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
994 for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
995 const int block = (row + idy) * 2 + (col + idx);
996 const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
997 uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
998 int16_t *const src_diff =
999 vp9_raster_block_offset_int16(BLOCK_8X8, block, p->src_diff);
1000 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
1001 xd->mi[0]->bmi[block].as_mode = mode;
1002 vp9_predict_intra_block(xd, 1, TX_4X4, mode,
1003 x->skip_encode ? src : dst,
1004 x->skip_encode ? src_stride : dst_stride, dst,
1005 dst_stride, col + idx, row + idy, 0);
1006 vpx_highbd_subtract_block(4, 4, src_diff, 8, src, src_stride, dst,
1007 dst_stride, xd->bd);
1008 if (xd->lossless) {
1009 const scan_order *so = &vp9_default_scan_orders[TX_4X4];
1010 const int coeff_ctx =
1011 combine_entropy_contexts(tempa[idx], templ[idy]);
1012 vp9_highbd_fwht4x4(src_diff, coeff, 8);
1013 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
1014 ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
1015 so->neighbors, cpi->sf.use_fast_coef_costing);
1016 tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0 ? 1 : 0);
1017 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
1018 goto next_highbd;
1019 vp9_highbd_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block), dst,
1020 dst_stride, p->eobs[block], xd->bd);
1021 } else {
1022 int64_t unused;
1023 const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
1024 const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
1025 const int coeff_ctx =
1026 combine_entropy_contexts(tempa[idx], templ[idy]);
1027 if (tx_type == DCT_DCT)
1028 vpx_highbd_fdct4x4(src_diff, coeff, 8);
1029 else
1030 vp9_highbd_fht4x4(src_diff, coeff, 8, tx_type);
1031 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
1032 ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
1033 so->neighbors, cpi->sf.use_fast_coef_costing);
1034 distortion += vp9_highbd_block_error_dispatch(
1035 coeff, BLOCK_OFFSET(pd->dqcoeff, block), 16,
1036 &unused, xd->bd) >>
1037 2;
1038 tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0 ? 1 : 0);
1039 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
1040 goto next_highbd;
1041 vp9_highbd_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block),
1042 dst, dst_stride, p->eobs[block], xd->bd);
1043 }
1044 }
1045 }
1046
1047 rate += ratey;
1048 this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
1049
1050 if (this_rd < best_rd) {
1051 *bestrate = rate;
1052 *bestratey = ratey;
1053 *bestdistortion = distortion;
1054 best_rd = this_rd;
1055 *best_mode = mode;
1056 memcpy(a, tempa, num_4x4_blocks_wide * sizeof(tempa[0]));
1057 memcpy(l, templ, num_4x4_blocks_high * sizeof(templ[0]));
1058 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
1059 memcpy(best_dst16 + idy * 8,
1060 CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
1061 num_4x4_blocks_wide * 4 * sizeof(uint16_t));
1062 }
1063 }
1064 next_highbd : {}
1065 }
1066 if (best_rd >= rd_thresh || x->skip_encode) return best_rd;
1067
1068 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
1069 memcpy(CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
1070 best_dst16 + idy * 8, num_4x4_blocks_wide * 4 * sizeof(uint16_t));
1071 }
1072
1073 return best_rd;
1074 }
1075 #endif // CONFIG_VP9_HIGHBITDEPTH
1076
1077 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
1078 int64_t this_rd;
1079 int ratey = 0;
1080 int64_t distortion = 0;
1081 int rate = bmode_costs[mode];
1082
1083 if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode))) continue;
1084
1085 // Only do the oblique modes if the best so far is
1086 // one of the neighboring directional modes
1087 if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
1088 if (conditional_skipintra(mode, *best_mode)) continue;
1089 }
1090
1091 memcpy(tempa, ta, num_4x4_blocks_wide * sizeof(ta[0]));
1092 memcpy(templ, tl, num_4x4_blocks_high * sizeof(tl[0]));
1093
1094 for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
1095 for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
1096 const int block = (row + idy) * 2 + (col + idx);
1097 const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
1098 uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
1099 int16_t *const src_diff =
1100 vp9_raster_block_offset_int16(BLOCK_8X8, block, p->src_diff);
1101 tran_low_t *const coeff = BLOCK_OFFSET(x->plane[0].coeff, block);
1102 xd->mi[0]->bmi[block].as_mode = mode;
1103 vp9_predict_intra_block(xd, 1, TX_4X4, mode, x->skip_encode ? src : dst,
1104 x->skip_encode ? src_stride : dst_stride, dst,
1105 dst_stride, col + idx, row + idy, 0);
1106 vpx_subtract_block(4, 4, src_diff, 8, src, src_stride, dst, dst_stride);
1107
1108 if (xd->lossless) {
1109 const scan_order *so = &vp9_default_scan_orders[TX_4X4];
1110 const int coeff_ctx =
1111 combine_entropy_contexts(tempa[idx], templ[idy]);
1112 vp9_fwht4x4(src_diff, coeff, 8);
1113 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
1114 ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
1115 so->neighbors, cpi->sf.use_fast_coef_costing);
1116 tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0) ? 1 : 0;
1117 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
1118 goto next;
1119 vp9_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block), dst, dst_stride,
1120 p->eobs[block]);
1121 } else {
1122 int64_t unused;
1123 const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
1124 const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
1125 const int coeff_ctx =
1126 combine_entropy_contexts(tempa[idx], templ[idy]);
1127 vp9_fht4x4(src_diff, coeff, 8, tx_type);
1128 vp9_regular_quantize_b_4x4(x, 0, block, so->scan, so->iscan);
1129 ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
1130 so->neighbors, cpi->sf.use_fast_coef_costing);
1131 tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0) ? 1 : 0;
1132 #if CONFIG_VP9_HIGHBITDEPTH
1133 distortion +=
1134 vp9_highbd_block_error_8bit(
1135 coeff, BLOCK_OFFSET(pd->dqcoeff, block), 16, &unused) >>
1136 2;
1137 #else
1138 distortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, block),
1139 16, &unused) >>
1140 2;
1141 #endif
1142 if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
1143 goto next;
1144 vp9_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block), dst,
1145 dst_stride, p->eobs[block]);
1146 }
1147 }
1148 }
1149
1150 rate += ratey;
1151 this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
1152
1153 if (this_rd < best_rd) {
1154 *bestrate = rate;
1155 *bestratey = ratey;
1156 *bestdistortion = distortion;
1157 best_rd = this_rd;
1158 *best_mode = mode;
1159 memcpy(a, tempa, num_4x4_blocks_wide * sizeof(tempa[0]));
1160 memcpy(l, templ, num_4x4_blocks_high * sizeof(templ[0]));
1161 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
1162 memcpy(best_dst + idy * 8, dst_init + idy * dst_stride,
1163 num_4x4_blocks_wide * 4);
1164 }
1165 next : {}
1166 }
1167
1168 if (best_rd >= rd_thresh || x->skip_encode) return best_rd;
1169
1170 for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
1171 memcpy(dst_init + idy * dst_stride, best_dst + idy * 8,
1172 num_4x4_blocks_wide * 4);
1173
1174 return best_rd;
1175 }
1176
rd_pick_intra_sub_8x8_y_mode(VP9_COMP * cpi,MACROBLOCK * mb,int * rate,int * rate_y,int64_t * distortion,int64_t best_rd)1177 static int64_t rd_pick_intra_sub_8x8_y_mode(VP9_COMP *cpi, MACROBLOCK *mb,
1178 int *rate, int *rate_y,
1179 int64_t *distortion,
1180 int64_t best_rd) {
1181 int i, j;
1182 const MACROBLOCKD *const xd = &mb->e_mbd;
1183 MODE_INFO *const mic = xd->mi[0];
1184 const MODE_INFO *above_mi = xd->above_mi;
1185 const MODE_INFO *left_mi = xd->left_mi;
1186 const BLOCK_SIZE bsize = xd->mi[0]->sb_type;
1187 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1188 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1189 int idx, idy;
1190 int cost = 0;
1191 int64_t total_distortion = 0;
1192 int tot_rate_y = 0;
1193 int64_t total_rd = 0;
1194 const int *bmode_costs = cpi->mbmode_cost;
1195
1196 // Pick modes for each sub-block (of size 4x4, 4x8, or 8x4) in an 8x8 block.
1197 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1198 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1199 PREDICTION_MODE best_mode = DC_PRED;
1200 int r = INT_MAX, ry = INT_MAX;
1201 int64_t d = INT64_MAX, this_rd = INT64_MAX;
1202 i = idy * 2 + idx;
1203 if (cpi->common.frame_type == KEY_FRAME) {
1204 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, i);
1205 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, i);
1206
1207 bmode_costs = cpi->y_mode_costs[A][L];
1208 }
1209
1210 this_rd = rd_pick_intra4x4block(
1211 cpi, mb, idy, idx, &best_mode, bmode_costs,
1212 xd->plane[0].above_context + idx, xd->plane[0].left_context + idy, &r,
1213 &ry, &d, bsize, best_rd - total_rd);
1214
1215 if (this_rd >= best_rd - total_rd) return INT64_MAX;
1216
1217 total_rd += this_rd;
1218 cost += r;
1219 total_distortion += d;
1220 tot_rate_y += ry;
1221
1222 mic->bmi[i].as_mode = best_mode;
1223 for (j = 1; j < num_4x4_blocks_high; ++j)
1224 mic->bmi[i + j * 2].as_mode = best_mode;
1225 for (j = 1; j < num_4x4_blocks_wide; ++j)
1226 mic->bmi[i + j].as_mode = best_mode;
1227
1228 if (total_rd >= best_rd) return INT64_MAX;
1229 }
1230 }
1231
1232 *rate = cost;
1233 *rate_y = tot_rate_y;
1234 *distortion = total_distortion;
1235 mic->mode = mic->bmi[3].as_mode;
1236
1237 return RDCOST(mb->rdmult, mb->rddiv, cost, total_distortion);
1238 }
1239
1240 // This function is used only for intra_only frames
rd_pick_intra_sby_mode(VP9_COMP * cpi,MACROBLOCK * x,int * rate,int * rate_tokenonly,int64_t * distortion,int * skippable,BLOCK_SIZE bsize,int64_t best_rd)1241 static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
1242 int *rate_tokenonly, int64_t *distortion,
1243 int *skippable, BLOCK_SIZE bsize,
1244 int64_t best_rd) {
1245 PREDICTION_MODE mode;
1246 PREDICTION_MODE mode_selected = DC_PRED;
1247 MACROBLOCKD *const xd = &x->e_mbd;
1248 MODE_INFO *const mic = xd->mi[0];
1249 int this_rate, this_rate_tokenonly, s;
1250 int64_t this_distortion, this_rd;
1251 TX_SIZE best_tx = TX_4X4;
1252 int *bmode_costs;
1253 const MODE_INFO *above_mi = xd->above_mi;
1254 const MODE_INFO *left_mi = xd->left_mi;
1255 const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
1256 const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
1257 bmode_costs = cpi->y_mode_costs[A][L];
1258
1259 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1260 /* Y Search for intra prediction mode */
1261 for (mode = DC_PRED; mode <= TM_PRED; mode++) {
1262 if (cpi->sf.use_nonrd_pick_mode) {
1263 // These speed features are turned on in hybrid non-RD and RD mode
1264 // for key frame coding in the context of real-time setting.
1265 if (conditional_skipintra(mode, mode_selected)) continue;
1266 if (*skippable) break;
1267 }
1268
1269 mic->mode = mode;
1270
1271 super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion, &s, NULL,
1272 bsize, best_rd);
1273
1274 if (this_rate_tokenonly == INT_MAX) continue;
1275
1276 this_rate = this_rate_tokenonly + bmode_costs[mode];
1277 this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1278
1279 if (this_rd < best_rd) {
1280 mode_selected = mode;
1281 best_rd = this_rd;
1282 best_tx = mic->tx_size;
1283 *rate = this_rate;
1284 *rate_tokenonly = this_rate_tokenonly;
1285 *distortion = this_distortion;
1286 *skippable = s;
1287 }
1288 }
1289
1290 mic->mode = mode_selected;
1291 mic->tx_size = best_tx;
1292
1293 return best_rd;
1294 }
1295
1296 // Return value 0: early termination triggered, no valid rd cost available;
1297 // 1: rd cost values are valid.
super_block_uvrd(const VP9_COMP * cpi,MACROBLOCK * x,int * rate,int64_t * distortion,int * skippable,int64_t * sse,BLOCK_SIZE bsize,int64_t ref_best_rd)1298 static int super_block_uvrd(const VP9_COMP *cpi, MACROBLOCK *x, int *rate,
1299 int64_t *distortion, int *skippable, int64_t *sse,
1300 BLOCK_SIZE bsize, int64_t ref_best_rd) {
1301 MACROBLOCKD *const xd = &x->e_mbd;
1302 MODE_INFO *const mi = xd->mi[0];
1303 const TX_SIZE uv_tx_size = get_uv_tx_size(mi, &xd->plane[1]);
1304 int plane;
1305 int pnrate = 0, pnskip = 1;
1306 int64_t pndist = 0, pnsse = 0;
1307 int is_cost_valid = 1;
1308
1309 if (ref_best_rd < 0) is_cost_valid = 0;
1310
1311 if (is_inter_block(mi) && is_cost_valid) {
1312 int plane;
1313 for (plane = 1; plane < MAX_MB_PLANE; ++plane)
1314 vp9_subtract_plane(x, bsize, plane);
1315 }
1316
1317 *rate = 0;
1318 *distortion = 0;
1319 *sse = 0;
1320 *skippable = 1;
1321
1322 for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
1323 txfm_rd_in_plane(cpi, x, &pnrate, &pndist, &pnskip, &pnsse, ref_best_rd,
1324 plane, bsize, uv_tx_size, cpi->sf.use_fast_coef_costing);
1325 if (pnrate == INT_MAX) {
1326 is_cost_valid = 0;
1327 break;
1328 }
1329 *rate += pnrate;
1330 *distortion += pndist;
1331 *sse += pnsse;
1332 *skippable &= pnskip;
1333 }
1334
1335 if (!is_cost_valid) {
1336 // reset cost value
1337 *rate = INT_MAX;
1338 *distortion = INT64_MAX;
1339 *sse = INT64_MAX;
1340 *skippable = 0;
1341 }
1342
1343 return is_cost_valid;
1344 }
1345
rd_pick_intra_sbuv_mode(VP9_COMP * cpi,MACROBLOCK * x,PICK_MODE_CONTEXT * ctx,int * rate,int * rate_tokenonly,int64_t * distortion,int * skippable,BLOCK_SIZE bsize,TX_SIZE max_tx_size)1346 static int64_t rd_pick_intra_sbuv_mode(VP9_COMP *cpi, MACROBLOCK *x,
1347 PICK_MODE_CONTEXT *ctx, int *rate,
1348 int *rate_tokenonly, int64_t *distortion,
1349 int *skippable, BLOCK_SIZE bsize,
1350 TX_SIZE max_tx_size) {
1351 MACROBLOCKD *xd = &x->e_mbd;
1352 PREDICTION_MODE mode;
1353 PREDICTION_MODE mode_selected = DC_PRED;
1354 int64_t best_rd = INT64_MAX, this_rd;
1355 int this_rate_tokenonly, this_rate, s;
1356 int64_t this_distortion, this_sse;
1357
1358 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1359 for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
1360 if (!(cpi->sf.intra_uv_mode_mask[max_tx_size] & (1 << mode))) continue;
1361 #if CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
1362 if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) &&
1363 (xd->above_mi == NULL || xd->left_mi == NULL) && need_top_left[mode])
1364 continue;
1365 #endif // CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
1366
1367 xd->mi[0]->uv_mode = mode;
1368
1369 if (!super_block_uvrd(cpi, x, &this_rate_tokenonly, &this_distortion, &s,
1370 &this_sse, bsize, best_rd))
1371 continue;
1372 this_rate =
1373 this_rate_tokenonly +
1374 cpi->intra_uv_mode_cost[cpi->common.frame_type][xd->mi[0]->mode][mode];
1375 this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1376
1377 if (this_rd < best_rd) {
1378 mode_selected = mode;
1379 best_rd = this_rd;
1380 *rate = this_rate;
1381 *rate_tokenonly = this_rate_tokenonly;
1382 *distortion = this_distortion;
1383 *skippable = s;
1384 if (!x->select_tx_size) swap_block_ptr(x, ctx, 2, 0, 1, MAX_MB_PLANE);
1385 }
1386 }
1387
1388 xd->mi[0]->uv_mode = mode_selected;
1389 return best_rd;
1390 }
1391
rd_sbuv_dcpred(const VP9_COMP * cpi,MACROBLOCK * x,int * rate,int * rate_tokenonly,int64_t * distortion,int * skippable,BLOCK_SIZE bsize)1392 static int64_t rd_sbuv_dcpred(const VP9_COMP *cpi, MACROBLOCK *x, int *rate,
1393 int *rate_tokenonly, int64_t *distortion,
1394 int *skippable, BLOCK_SIZE bsize) {
1395 const VP9_COMMON *cm = &cpi->common;
1396 int64_t unused;
1397
1398 x->e_mbd.mi[0]->uv_mode = DC_PRED;
1399 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1400 super_block_uvrd(cpi, x, rate_tokenonly, distortion, skippable, &unused,
1401 bsize, INT64_MAX);
1402 *rate =
1403 *rate_tokenonly +
1404 cpi->intra_uv_mode_cost[cm->frame_type][x->e_mbd.mi[0]->mode][DC_PRED];
1405 return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
1406 }
1407
choose_intra_uv_mode(VP9_COMP * cpi,MACROBLOCK * const x,PICK_MODE_CONTEXT * ctx,BLOCK_SIZE bsize,TX_SIZE max_tx_size,int * rate_uv,int * rate_uv_tokenonly,int64_t * dist_uv,int * skip_uv,PREDICTION_MODE * mode_uv)1408 static void choose_intra_uv_mode(VP9_COMP *cpi, MACROBLOCK *const x,
1409 PICK_MODE_CONTEXT *ctx, BLOCK_SIZE bsize,
1410 TX_SIZE max_tx_size, int *rate_uv,
1411 int *rate_uv_tokenonly, int64_t *dist_uv,
1412 int *skip_uv, PREDICTION_MODE *mode_uv) {
1413 // Use an estimated rd for uv_intra based on DC_PRED if the
1414 // appropriate speed flag is set.
1415 if (cpi->sf.use_uv_intra_rd_estimate) {
1416 rd_sbuv_dcpred(cpi, x, rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
1417 bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize);
1418 // Else do a proper rd search for each possible transform size that may
1419 // be considered in the main rd loop.
1420 } else {
1421 rd_pick_intra_sbuv_mode(cpi, x, ctx, rate_uv, rate_uv_tokenonly, dist_uv,
1422 skip_uv, bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize,
1423 max_tx_size);
1424 }
1425 *mode_uv = x->e_mbd.mi[0]->uv_mode;
1426 }
1427
cost_mv_ref(const VP9_COMP * cpi,PREDICTION_MODE mode,int mode_context)1428 static int cost_mv_ref(const VP9_COMP *cpi, PREDICTION_MODE mode,
1429 int mode_context) {
1430 assert(is_inter_mode(mode));
1431 return cpi->inter_mode_cost[mode_context][INTER_OFFSET(mode)];
1432 }
1433
set_and_cost_bmi_mvs(VP9_COMP * cpi,MACROBLOCK * x,MACROBLOCKD * xd,int i,PREDICTION_MODE mode,int_mv this_mv[2],int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],int_mv seg_mvs[MAX_REF_FRAMES],int_mv * best_ref_mv[2],const int * mvjcost,int * mvcost[2])1434 static int set_and_cost_bmi_mvs(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
1435 int i, PREDICTION_MODE mode, int_mv this_mv[2],
1436 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1437 int_mv seg_mvs[MAX_REF_FRAMES],
1438 int_mv *best_ref_mv[2], const int *mvjcost,
1439 int *mvcost[2]) {
1440 MODE_INFO *const mi = xd->mi[0];
1441 const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1442 int thismvcost = 0;
1443 int idx, idy;
1444 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[mi->sb_type];
1445 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mi->sb_type];
1446 const int is_compound = has_second_ref(mi);
1447
1448 switch (mode) {
1449 case NEWMV:
1450 this_mv[0].as_int = seg_mvs[mi->ref_frame[0]].as_int;
1451 thismvcost += vp9_mv_bit_cost(&this_mv[0].as_mv, &best_ref_mv[0]->as_mv,
1452 mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1453 if (is_compound) {
1454 this_mv[1].as_int = seg_mvs[mi->ref_frame[1]].as_int;
1455 thismvcost += vp9_mv_bit_cost(&this_mv[1].as_mv, &best_ref_mv[1]->as_mv,
1456 mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1457 }
1458 break;
1459 case NEARMV:
1460 case NEARESTMV:
1461 this_mv[0].as_int = frame_mv[mode][mi->ref_frame[0]].as_int;
1462 if (is_compound)
1463 this_mv[1].as_int = frame_mv[mode][mi->ref_frame[1]].as_int;
1464 break;
1465 case ZEROMV:
1466 this_mv[0].as_int = 0;
1467 if (is_compound) this_mv[1].as_int = 0;
1468 break;
1469 default: break;
1470 }
1471
1472 mi->bmi[i].as_mv[0].as_int = this_mv[0].as_int;
1473 if (is_compound) mi->bmi[i].as_mv[1].as_int = this_mv[1].as_int;
1474
1475 mi->bmi[i].as_mode = mode;
1476
1477 for (idy = 0; idy < num_4x4_blocks_high; ++idy)
1478 for (idx = 0; idx < num_4x4_blocks_wide; ++idx)
1479 memmove(&mi->bmi[i + idy * 2 + idx], &mi->bmi[i], sizeof(mi->bmi[i]));
1480
1481 return cost_mv_ref(cpi, mode, mbmi_ext->mode_context[mi->ref_frame[0]]) +
1482 thismvcost;
1483 }
1484
encode_inter_mb_segment(VP9_COMP * cpi,MACROBLOCK * x,int64_t best_yrd,int i,int * labelyrate,int64_t * distortion,int64_t * sse,ENTROPY_CONTEXT * ta,ENTROPY_CONTEXT * tl,int mi_row,int mi_col)1485 static int64_t encode_inter_mb_segment(VP9_COMP *cpi, MACROBLOCK *x,
1486 int64_t best_yrd, int i, int *labelyrate,
1487 int64_t *distortion, int64_t *sse,
1488 ENTROPY_CONTEXT *ta, ENTROPY_CONTEXT *tl,
1489 int mi_row, int mi_col) {
1490 int k;
1491 MACROBLOCKD *xd = &x->e_mbd;
1492 struct macroblockd_plane *const pd = &xd->plane[0];
1493 struct macroblock_plane *const p = &x->plane[0];
1494 MODE_INFO *const mi = xd->mi[0];
1495 const BLOCK_SIZE plane_bsize = get_plane_block_size(mi->sb_type, pd);
1496 const int width = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
1497 const int height = 4 * num_4x4_blocks_high_lookup[plane_bsize];
1498 int idx, idy;
1499
1500 const uint8_t *const src =
1501 &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
1502 uint8_t *const dst =
1503 &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)];
1504 int64_t thisdistortion = 0, thissse = 0;
1505 int thisrate = 0, ref;
1506 const scan_order *so = &vp9_default_scan_orders[TX_4X4];
1507 const int is_compound = has_second_ref(mi);
1508 const InterpKernel *kernel = vp9_filter_kernels[mi->interp_filter];
1509
1510 for (ref = 0; ref < 1 + is_compound; ++ref) {
1511 const int bw = b_width_log2_lookup[BLOCK_8X8];
1512 const int h = 4 * (i >> bw);
1513 const int w = 4 * (i & ((1 << bw) - 1));
1514 const struct scale_factors *sf = &xd->block_refs[ref]->sf;
1515 int y_stride = pd->pre[ref].stride;
1516 uint8_t *pre = pd->pre[ref].buf + (h * pd->pre[ref].stride + w);
1517
1518 if (vp9_is_scaled(sf)) {
1519 const int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
1520 const int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
1521
1522 y_stride = xd->block_refs[ref]->buf->y_stride;
1523 pre = xd->block_refs[ref]->buf->y_buffer;
1524 pre += scaled_buffer_offset(x_start + w, y_start + h, y_stride, sf);
1525 }
1526 #if CONFIG_VP9_HIGHBITDEPTH
1527 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1528 vp9_highbd_build_inter_predictor(
1529 pre, y_stride, dst, pd->dst.stride, &mi->bmi[i].as_mv[ref].as_mv,
1530 &xd->block_refs[ref]->sf, width, height, ref, kernel, MV_PRECISION_Q3,
1531 mi_col * MI_SIZE + 4 * (i % 2), mi_row * MI_SIZE + 4 * (i / 2),
1532 xd->bd);
1533 } else {
1534 vp9_build_inter_predictor(
1535 pre, y_stride, dst, pd->dst.stride, &mi->bmi[i].as_mv[ref].as_mv,
1536 &xd->block_refs[ref]->sf, width, height, ref, kernel, MV_PRECISION_Q3,
1537 mi_col * MI_SIZE + 4 * (i % 2), mi_row * MI_SIZE + 4 * (i / 2));
1538 }
1539 #else
1540 vp9_build_inter_predictor(
1541 pre, y_stride, dst, pd->dst.stride, &mi->bmi[i].as_mv[ref].as_mv,
1542 &xd->block_refs[ref]->sf, width, height, ref, kernel, MV_PRECISION_Q3,
1543 mi_col * MI_SIZE + 4 * (i % 2), mi_row * MI_SIZE + 4 * (i / 2));
1544 #endif // CONFIG_VP9_HIGHBITDEPTH
1545 }
1546
1547 #if CONFIG_VP9_HIGHBITDEPTH
1548 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1549 vpx_highbd_subtract_block(
1550 height, width, vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1551 8, src, p->src.stride, dst, pd->dst.stride, xd->bd);
1552 } else {
1553 vpx_subtract_block(height, width,
1554 vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1555 8, src, p->src.stride, dst, pd->dst.stride);
1556 }
1557 #else
1558 vpx_subtract_block(height, width,
1559 vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1560 8, src, p->src.stride, dst, pd->dst.stride);
1561 #endif // CONFIG_VP9_HIGHBITDEPTH
1562
1563 k = i;
1564 for (idy = 0; idy < height / 4; ++idy) {
1565 for (idx = 0; idx < width / 4; ++idx) {
1566 #if CONFIG_VP9_HIGHBITDEPTH
1567 const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
1568 #endif
1569 int64_t ssz, rd, rd1, rd2;
1570 tran_low_t *coeff;
1571 int coeff_ctx;
1572 k += (idy * 2 + idx);
1573 coeff_ctx = combine_entropy_contexts(ta[k & 1], tl[k >> 1]);
1574 coeff = BLOCK_OFFSET(p->coeff, k);
1575 x->fwd_txm4x4(vp9_raster_block_offset_int16(BLOCK_8X8, k, p->src_diff),
1576 coeff, 8);
1577 vp9_regular_quantize_b_4x4(x, 0, k, so->scan, so->iscan);
1578 #if CONFIG_VP9_HIGHBITDEPTH
1579 thisdistortion += vp9_highbd_block_error_dispatch(
1580 coeff, BLOCK_OFFSET(pd->dqcoeff, k), 16, &ssz, bd);
1581 #else
1582 thisdistortion +=
1583 vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k), 16, &ssz);
1584 #endif // CONFIG_VP9_HIGHBITDEPTH
1585 thissse += ssz;
1586 thisrate += cost_coeffs(x, 0, k, TX_4X4, coeff_ctx, so->scan,
1587 so->neighbors, cpi->sf.use_fast_coef_costing);
1588 ta[k & 1] = tl[k >> 1] = (x->plane[0].eobs[k] > 0) ? 1 : 0;
1589 rd1 = RDCOST(x->rdmult, x->rddiv, thisrate, thisdistortion >> 2);
1590 rd2 = RDCOST(x->rdmult, x->rddiv, 0, thissse >> 2);
1591 rd = VPXMIN(rd1, rd2);
1592 if (rd >= best_yrd) return INT64_MAX;
1593 }
1594 }
1595
1596 *distortion = thisdistortion >> 2;
1597 *labelyrate = thisrate;
1598 *sse = thissse >> 2;
1599
1600 return RDCOST(x->rdmult, x->rddiv, *labelyrate, *distortion);
1601 }
1602
1603 typedef struct {
1604 int eobs;
1605 int brate;
1606 int byrate;
1607 int64_t bdist;
1608 int64_t bsse;
1609 int64_t brdcost;
1610 int_mv mvs[2];
1611 ENTROPY_CONTEXT ta[2];
1612 ENTROPY_CONTEXT tl[2];
1613 } SEG_RDSTAT;
1614
1615 typedef struct {
1616 int_mv *ref_mv[2];
1617 int_mv mvp;
1618
1619 int64_t segment_rd;
1620 int r;
1621 int64_t d;
1622 int64_t sse;
1623 int segment_yrate;
1624 PREDICTION_MODE modes[4];
1625 SEG_RDSTAT rdstat[4][INTER_MODES];
1626 int mvthresh;
1627 } BEST_SEG_INFO;
1628
mv_check_bounds(const MvLimits * mv_limits,const MV * mv)1629 static INLINE int mv_check_bounds(const MvLimits *mv_limits, const MV *mv) {
1630 return (mv->row >> 3) < mv_limits->row_min ||
1631 (mv->row >> 3) > mv_limits->row_max ||
1632 (mv->col >> 3) < mv_limits->col_min ||
1633 (mv->col >> 3) > mv_limits->col_max;
1634 }
1635
mi_buf_shift(MACROBLOCK * x,int i)1636 static INLINE void mi_buf_shift(MACROBLOCK *x, int i) {
1637 MODE_INFO *const mi = x->e_mbd.mi[0];
1638 struct macroblock_plane *const p = &x->plane[0];
1639 struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
1640
1641 p->src.buf =
1642 &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
1643 assert(((intptr_t)pd->pre[0].buf & 0x7) == 0);
1644 pd->pre[0].buf =
1645 &pd->pre[0].buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->pre[0].stride)];
1646 if (has_second_ref(mi))
1647 pd->pre[1].buf =
1648 &pd->pre[1]
1649 .buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->pre[1].stride)];
1650 }
1651
mi_buf_restore(MACROBLOCK * x,struct buf_2d orig_src,struct buf_2d orig_pre[2])1652 static INLINE void mi_buf_restore(MACROBLOCK *x, struct buf_2d orig_src,
1653 struct buf_2d orig_pre[2]) {
1654 MODE_INFO *mi = x->e_mbd.mi[0];
1655 x->plane[0].src = orig_src;
1656 x->e_mbd.plane[0].pre[0] = orig_pre[0];
1657 if (has_second_ref(mi)) x->e_mbd.plane[0].pre[1] = orig_pre[1];
1658 }
1659
mv_has_subpel(const MV * mv)1660 static INLINE int mv_has_subpel(const MV *mv) {
1661 return (mv->row & 0x0F) || (mv->col & 0x0F);
1662 }
1663
1664 // Check if NEARESTMV/NEARMV/ZEROMV is the cheapest way encode zero motion.
1665 // TODO(aconverse): Find out if this is still productive then clean up or remove
check_best_zero_mv(const VP9_COMP * cpi,const uint8_t mode_context[MAX_REF_FRAMES],int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],int this_mode,const MV_REFERENCE_FRAME ref_frames[2])1666 static int check_best_zero_mv(const VP9_COMP *cpi,
1667 const uint8_t mode_context[MAX_REF_FRAMES],
1668 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1669 int this_mode,
1670 const MV_REFERENCE_FRAME ref_frames[2]) {
1671 if ((this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) &&
1672 frame_mv[this_mode][ref_frames[0]].as_int == 0 &&
1673 (ref_frames[1] == NONE ||
1674 frame_mv[this_mode][ref_frames[1]].as_int == 0)) {
1675 int rfc = mode_context[ref_frames[0]];
1676 int c1 = cost_mv_ref(cpi, NEARMV, rfc);
1677 int c2 = cost_mv_ref(cpi, NEARESTMV, rfc);
1678 int c3 = cost_mv_ref(cpi, ZEROMV, rfc);
1679
1680 if (this_mode == NEARMV) {
1681 if (c1 > c3) return 0;
1682 } else if (this_mode == NEARESTMV) {
1683 if (c2 > c3) return 0;
1684 } else {
1685 assert(this_mode == ZEROMV);
1686 if (ref_frames[1] == NONE) {
1687 if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0) ||
1688 (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0))
1689 return 0;
1690 } else {
1691 if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0 &&
1692 frame_mv[NEARESTMV][ref_frames[1]].as_int == 0) ||
1693 (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0 &&
1694 frame_mv[NEARMV][ref_frames[1]].as_int == 0))
1695 return 0;
1696 }
1697 }
1698 }
1699 return 1;
1700 }
1701
joint_motion_search(VP9_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,int_mv * frame_mv,int mi_row,int mi_col,int_mv single_newmv[MAX_REF_FRAMES],int * rate_mv)1702 static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
1703 int_mv *frame_mv, int mi_row, int mi_col,
1704 int_mv single_newmv[MAX_REF_FRAMES],
1705 int *rate_mv) {
1706 const VP9_COMMON *const cm = &cpi->common;
1707 const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
1708 const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
1709 MACROBLOCKD *xd = &x->e_mbd;
1710 MODE_INFO *mi = xd->mi[0];
1711 const int refs[2] = { mi->ref_frame[0],
1712 mi->ref_frame[1] < 0 ? 0 : mi->ref_frame[1] };
1713 int_mv ref_mv[2];
1714 int ite, ref;
1715 const InterpKernel *kernel = vp9_filter_kernels[mi->interp_filter];
1716 struct scale_factors sf;
1717
1718 // Do joint motion search in compound mode to get more accurate mv.
1719 struct buf_2d backup_yv12[2][MAX_MB_PLANE];
1720 uint32_t last_besterr[2] = { UINT_MAX, UINT_MAX };
1721 const YV12_BUFFER_CONFIG *const scaled_ref_frame[2] = {
1722 vp9_get_scaled_ref_frame(cpi, mi->ref_frame[0]),
1723 vp9_get_scaled_ref_frame(cpi, mi->ref_frame[1])
1724 };
1725
1726 // Prediction buffer from second frame.
1727 #if CONFIG_VP9_HIGHBITDEPTH
1728 DECLARE_ALIGNED(16, uint16_t, second_pred_alloc_16[64 * 64]);
1729 uint8_t *second_pred;
1730 #else
1731 DECLARE_ALIGNED(16, uint8_t, second_pred[64 * 64]);
1732 #endif // CONFIG_VP9_HIGHBITDEPTH
1733
1734 for (ref = 0; ref < 2; ++ref) {
1735 ref_mv[ref] = x->mbmi_ext->ref_mvs[refs[ref]][0];
1736
1737 if (scaled_ref_frame[ref]) {
1738 int i;
1739 // Swap out the reference frame for a version that's been scaled to
1740 // match the resolution of the current frame, allowing the existing
1741 // motion search code to be used without additional modifications.
1742 for (i = 0; i < MAX_MB_PLANE; i++)
1743 backup_yv12[ref][i] = xd->plane[i].pre[ref];
1744 vp9_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
1745 NULL);
1746 }
1747
1748 frame_mv[refs[ref]].as_int = single_newmv[refs[ref]].as_int;
1749 }
1750
1751 // Since we have scaled the reference frames to match the size of the current
1752 // frame we must use a unit scaling factor during mode selection.
1753 #if CONFIG_VP9_HIGHBITDEPTH
1754 vp9_setup_scale_factors_for_frame(&sf, cm->width, cm->height, cm->width,
1755 cm->height, cm->use_highbitdepth);
1756 #else
1757 vp9_setup_scale_factors_for_frame(&sf, cm->width, cm->height, cm->width,
1758 cm->height);
1759 #endif // CONFIG_VP9_HIGHBITDEPTH
1760
1761 // Allow joint search multiple times iteratively for each reference frame
1762 // and break out of the search loop if it couldn't find a better mv.
1763 for (ite = 0; ite < 4; ite++) {
1764 struct buf_2d ref_yv12[2];
1765 uint32_t bestsme = UINT_MAX;
1766 int sadpb = x->sadperbit16;
1767 MV tmp_mv;
1768 int search_range = 3;
1769
1770 const MvLimits tmp_mv_limits = x->mv_limits;
1771 int id = ite % 2; // Even iterations search in the first reference frame,
1772 // odd iterations search in the second. The predictor
1773 // found for the 'other' reference frame is factored in.
1774
1775 // Initialized here because of compiler problem in Visual Studio.
1776 ref_yv12[0] = xd->plane[0].pre[0];
1777 ref_yv12[1] = xd->plane[0].pre[1];
1778
1779 // Get the prediction block from the 'other' reference frame.
1780 #if CONFIG_VP9_HIGHBITDEPTH
1781 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1782 second_pred = CONVERT_TO_BYTEPTR(second_pred_alloc_16);
1783 vp9_highbd_build_inter_predictor(
1784 ref_yv12[!id].buf, ref_yv12[!id].stride, second_pred, pw,
1785 &frame_mv[refs[!id]].as_mv, &sf, pw, ph, 0, kernel, MV_PRECISION_Q3,
1786 mi_col * MI_SIZE, mi_row * MI_SIZE, xd->bd);
1787 } else {
1788 second_pred = (uint8_t *)second_pred_alloc_16;
1789 vp9_build_inter_predictor(ref_yv12[!id].buf, ref_yv12[!id].stride,
1790 second_pred, pw, &frame_mv[refs[!id]].as_mv,
1791 &sf, pw, ph, 0, kernel, MV_PRECISION_Q3,
1792 mi_col * MI_SIZE, mi_row * MI_SIZE);
1793 }
1794 #else
1795 vp9_build_inter_predictor(ref_yv12[!id].buf, ref_yv12[!id].stride,
1796 second_pred, pw, &frame_mv[refs[!id]].as_mv, &sf,
1797 pw, ph, 0, kernel, MV_PRECISION_Q3,
1798 mi_col * MI_SIZE, mi_row * MI_SIZE);
1799 #endif // CONFIG_VP9_HIGHBITDEPTH
1800
1801 // Do compound motion search on the current reference frame.
1802 if (id) xd->plane[0].pre[0] = ref_yv12[id];
1803 vp9_set_mv_search_range(&x->mv_limits, &ref_mv[id].as_mv);
1804
1805 // Use the mv result from the single mode as mv predictor.
1806 tmp_mv = frame_mv[refs[id]].as_mv;
1807
1808 tmp_mv.col >>= 3;
1809 tmp_mv.row >>= 3;
1810
1811 // Small-range full-pixel motion search.
1812 bestsme = vp9_refining_search_8p_c(x, &tmp_mv, sadpb, search_range,
1813 &cpi->fn_ptr[bsize], &ref_mv[id].as_mv,
1814 second_pred);
1815 if (bestsme < UINT_MAX)
1816 bestsme = vp9_get_mvpred_av_var(x, &tmp_mv, &ref_mv[id].as_mv,
1817 second_pred, &cpi->fn_ptr[bsize], 1);
1818
1819 x->mv_limits = tmp_mv_limits;
1820
1821 if (bestsme < UINT_MAX) {
1822 uint32_t dis; /* TODO: use dis in distortion calculation later. */
1823 uint32_t sse;
1824 bestsme = cpi->find_fractional_mv_step(
1825 x, &tmp_mv, &ref_mv[id].as_mv, cpi->common.allow_high_precision_mv,
1826 x->errorperbit, &cpi->fn_ptr[bsize], 0,
1827 cpi->sf.mv.subpel_iters_per_step, NULL, x->nmvjointcost, x->mvcost,
1828 &dis, &sse, second_pred, pw, ph);
1829 }
1830
1831 // Restore the pointer to the first (possibly scaled) prediction buffer.
1832 if (id) xd->plane[0].pre[0] = ref_yv12[0];
1833
1834 if (bestsme < last_besterr[id]) {
1835 frame_mv[refs[id]].as_mv = tmp_mv;
1836 last_besterr[id] = bestsme;
1837 } else {
1838 break;
1839 }
1840 }
1841
1842 *rate_mv = 0;
1843
1844 for (ref = 0; ref < 2; ++ref) {
1845 if (scaled_ref_frame[ref]) {
1846 // Restore the prediction frame pointers to their unscaled versions.
1847 int i;
1848 for (i = 0; i < MAX_MB_PLANE; i++)
1849 xd->plane[i].pre[ref] = backup_yv12[ref][i];
1850 }
1851
1852 *rate_mv += vp9_mv_bit_cost(&frame_mv[refs[ref]].as_mv,
1853 &x->mbmi_ext->ref_mvs[refs[ref]][0].as_mv,
1854 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1855 }
1856 }
1857
rd_pick_best_sub8x8_mode(VP9_COMP * cpi,MACROBLOCK * x,int_mv * best_ref_mv,int_mv * second_best_ref_mv,int64_t best_rd,int * returntotrate,int * returnyrate,int64_t * returndistortion,int * skippable,int64_t * psse,int mvthresh,int_mv seg_mvs[4][MAX_REF_FRAMES],BEST_SEG_INFO * bsi_buf,int filter_idx,int mi_row,int mi_col)1858 static int64_t rd_pick_best_sub8x8_mode(
1859 VP9_COMP *cpi, MACROBLOCK *x, int_mv *best_ref_mv,
1860 int_mv *second_best_ref_mv, int64_t best_rd, int *returntotrate,
1861 int *returnyrate, int64_t *returndistortion, int *skippable, int64_t *psse,
1862 int mvthresh, int_mv seg_mvs[4][MAX_REF_FRAMES], BEST_SEG_INFO *bsi_buf,
1863 int filter_idx, int mi_row, int mi_col) {
1864 int i;
1865 BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
1866 MACROBLOCKD *xd = &x->e_mbd;
1867 MODE_INFO *mi = xd->mi[0];
1868 int mode_idx;
1869 int k, br = 0, idx, idy;
1870 int64_t bd = 0, block_sse = 0;
1871 PREDICTION_MODE this_mode;
1872 VP9_COMMON *cm = &cpi->common;
1873 struct macroblock_plane *const p = &x->plane[0];
1874 struct macroblockd_plane *const pd = &xd->plane[0];
1875 const int label_count = 4;
1876 int64_t this_segment_rd = 0;
1877 int label_mv_thresh;
1878 int segmentyrate = 0;
1879 const BLOCK_SIZE bsize = mi->sb_type;
1880 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1881 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1882 ENTROPY_CONTEXT t_above[2], t_left[2];
1883 int subpelmv = 1, have_ref = 0;
1884 SPEED_FEATURES *const sf = &cpi->sf;
1885 const int has_second_rf = has_second_ref(mi);
1886 const int inter_mode_mask = sf->inter_mode_mask[bsize];
1887 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1888
1889 vp9_zero(*bsi);
1890
1891 bsi->segment_rd = best_rd;
1892 bsi->ref_mv[0] = best_ref_mv;
1893 bsi->ref_mv[1] = second_best_ref_mv;
1894 bsi->mvp.as_int = best_ref_mv->as_int;
1895 bsi->mvthresh = mvthresh;
1896
1897 for (i = 0; i < 4; i++) bsi->modes[i] = ZEROMV;
1898
1899 memcpy(t_above, pd->above_context, sizeof(t_above));
1900 memcpy(t_left, pd->left_context, sizeof(t_left));
1901
1902 // 64 makes this threshold really big effectively
1903 // making it so that we very rarely check mvs on
1904 // segments. setting this to 1 would make mv thresh
1905 // roughly equal to what it is for macroblocks
1906 label_mv_thresh = 1 * bsi->mvthresh / label_count;
1907
1908 // Segmentation method overheads
1909 for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1910 for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1911 // TODO(jingning,rbultje): rewrite the rate-distortion optimization
1912 // loop for 4x4/4x8/8x4 block coding. to be replaced with new rd loop
1913 int_mv mode_mv[MB_MODE_COUNT][2];
1914 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1915 PREDICTION_MODE mode_selected = ZEROMV;
1916 int64_t best_rd = INT64_MAX;
1917 const int i = idy * 2 + idx;
1918 int ref;
1919
1920 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
1921 const MV_REFERENCE_FRAME frame = mi->ref_frame[ref];
1922 frame_mv[ZEROMV][frame].as_int = 0;
1923 vp9_append_sub8x8_mvs_for_idx(
1924 cm, xd, i, ref, mi_row, mi_col, &frame_mv[NEARESTMV][frame],
1925 &frame_mv[NEARMV][frame], mbmi_ext->mode_context);
1926 }
1927
1928 // search for the best motion vector on this segment
1929 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
1930 const struct buf_2d orig_src = x->plane[0].src;
1931 struct buf_2d orig_pre[2];
1932
1933 mode_idx = INTER_OFFSET(this_mode);
1934 bsi->rdstat[i][mode_idx].brdcost = INT64_MAX;
1935 if (!(inter_mode_mask & (1 << this_mode))) continue;
1936
1937 if (!check_best_zero_mv(cpi, mbmi_ext->mode_context, frame_mv,
1938 this_mode, mi->ref_frame))
1939 continue;
1940
1941 memcpy(orig_pre, pd->pre, sizeof(orig_pre));
1942 memcpy(bsi->rdstat[i][mode_idx].ta, t_above,
1943 sizeof(bsi->rdstat[i][mode_idx].ta));
1944 memcpy(bsi->rdstat[i][mode_idx].tl, t_left,
1945 sizeof(bsi->rdstat[i][mode_idx].tl));
1946
1947 // motion search for newmv (single predictor case only)
1948 if (!has_second_rf && this_mode == NEWMV &&
1949 seg_mvs[i][mi->ref_frame[0]].as_int == INVALID_MV) {
1950 MV *const new_mv = &mode_mv[NEWMV][0].as_mv;
1951 int step_param = 0;
1952 uint32_t bestsme = UINT_MAX;
1953 int sadpb = x->sadperbit4;
1954 MV mvp_full;
1955 int max_mv;
1956 int cost_list[5];
1957 const MvLimits tmp_mv_limits = x->mv_limits;
1958
1959 /* Is the best so far sufficiently good that we cant justify doing
1960 * and new motion search. */
1961 if (best_rd < label_mv_thresh) break;
1962
1963 if (cpi->oxcf.mode != BEST) {
1964 // use previous block's result as next block's MV predictor.
1965 if (i > 0) {
1966 bsi->mvp.as_int = mi->bmi[i - 1].as_mv[0].as_int;
1967 if (i == 2) bsi->mvp.as_int = mi->bmi[i - 2].as_mv[0].as_int;
1968 }
1969 }
1970 if (i == 0)
1971 max_mv = x->max_mv_context[mi->ref_frame[0]];
1972 else
1973 max_mv =
1974 VPXMAX(abs(bsi->mvp.as_mv.row), abs(bsi->mvp.as_mv.col)) >> 3;
1975
1976 if (sf->mv.auto_mv_step_size && cm->show_frame) {
1977 // Take wtd average of the step_params based on the last frame's
1978 // max mv magnitude and the best ref mvs of the current block for
1979 // the given reference.
1980 step_param =
1981 (vp9_init_search_range(max_mv) + cpi->mv_step_param) / 2;
1982 } else {
1983 step_param = cpi->mv_step_param;
1984 }
1985
1986 mvp_full.row = bsi->mvp.as_mv.row >> 3;
1987 mvp_full.col = bsi->mvp.as_mv.col >> 3;
1988
1989 if (sf->adaptive_motion_search) {
1990 mvp_full.row = x->pred_mv[mi->ref_frame[0]].row >> 3;
1991 mvp_full.col = x->pred_mv[mi->ref_frame[0]].col >> 3;
1992 step_param = VPXMAX(step_param, 8);
1993 }
1994
1995 // adjust src pointer for this block
1996 mi_buf_shift(x, i);
1997
1998 vp9_set_mv_search_range(&x->mv_limits, &bsi->ref_mv[0]->as_mv);
1999
2000 bestsme = vp9_full_pixel_search(
2001 cpi, x, bsize, &mvp_full, step_param, sadpb,
2002 sf->mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL,
2003 &bsi->ref_mv[0]->as_mv, new_mv, INT_MAX, 1);
2004
2005 x->mv_limits = tmp_mv_limits;
2006
2007 if (bestsme < UINT_MAX) {
2008 uint32_t distortion;
2009 cpi->find_fractional_mv_step(
2010 x, new_mv, &bsi->ref_mv[0]->as_mv, cm->allow_high_precision_mv,
2011 x->errorperbit, &cpi->fn_ptr[bsize], sf->mv.subpel_force_stop,
2012 sf->mv.subpel_iters_per_step, cond_cost_list(cpi, cost_list),
2013 x->nmvjointcost, x->mvcost, &distortion,
2014 &x->pred_sse[mi->ref_frame[0]], NULL, 0, 0);
2015
2016 // save motion search result for use in compound prediction
2017 seg_mvs[i][mi->ref_frame[0]].as_mv = *new_mv;
2018 }
2019
2020 if (sf->adaptive_motion_search)
2021 x->pred_mv[mi->ref_frame[0]] = *new_mv;
2022
2023 // restore src pointers
2024 mi_buf_restore(x, orig_src, orig_pre);
2025 }
2026
2027 if (has_second_rf) {
2028 if (seg_mvs[i][mi->ref_frame[1]].as_int == INVALID_MV ||
2029 seg_mvs[i][mi->ref_frame[0]].as_int == INVALID_MV)
2030 continue;
2031 }
2032
2033 if (has_second_rf && this_mode == NEWMV &&
2034 mi->interp_filter == EIGHTTAP) {
2035 // adjust src pointers
2036 mi_buf_shift(x, i);
2037 if (sf->comp_inter_joint_search_thresh <= bsize) {
2038 int rate_mv;
2039 joint_motion_search(cpi, x, bsize, frame_mv[this_mode], mi_row,
2040 mi_col, seg_mvs[i], &rate_mv);
2041 seg_mvs[i][mi->ref_frame[0]].as_int =
2042 frame_mv[this_mode][mi->ref_frame[0]].as_int;
2043 seg_mvs[i][mi->ref_frame[1]].as_int =
2044 frame_mv[this_mode][mi->ref_frame[1]].as_int;
2045 }
2046 // restore src pointers
2047 mi_buf_restore(x, orig_src, orig_pre);
2048 }
2049
2050 bsi->rdstat[i][mode_idx].brate = set_and_cost_bmi_mvs(
2051 cpi, x, xd, i, this_mode, mode_mv[this_mode], frame_mv, seg_mvs[i],
2052 bsi->ref_mv, x->nmvjointcost, x->mvcost);
2053
2054 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
2055 bsi->rdstat[i][mode_idx].mvs[ref].as_int =
2056 mode_mv[this_mode][ref].as_int;
2057 if (num_4x4_blocks_wide > 1)
2058 bsi->rdstat[i + 1][mode_idx].mvs[ref].as_int =
2059 mode_mv[this_mode][ref].as_int;
2060 if (num_4x4_blocks_high > 1)
2061 bsi->rdstat[i + 2][mode_idx].mvs[ref].as_int =
2062 mode_mv[this_mode][ref].as_int;
2063 }
2064
2065 // Trap vectors that reach beyond the UMV borders
2066 if (mv_check_bounds(&x->mv_limits, &mode_mv[this_mode][0].as_mv) ||
2067 (has_second_rf &&
2068 mv_check_bounds(&x->mv_limits, &mode_mv[this_mode][1].as_mv)))
2069 continue;
2070
2071 if (filter_idx > 0) {
2072 BEST_SEG_INFO *ref_bsi = bsi_buf;
2073 subpelmv = 0;
2074 have_ref = 1;
2075
2076 for (ref = 0; ref < 1 + has_second_rf; ++ref) {
2077 subpelmv |= mv_has_subpel(&mode_mv[this_mode][ref].as_mv);
2078 have_ref &= mode_mv[this_mode][ref].as_int ==
2079 ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
2080 }
2081
2082 if (filter_idx > 1 && !subpelmv && !have_ref) {
2083 ref_bsi = bsi_buf + 1;
2084 have_ref = 1;
2085 for (ref = 0; ref < 1 + has_second_rf; ++ref)
2086 have_ref &= mode_mv[this_mode][ref].as_int ==
2087 ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
2088 }
2089
2090 if (!subpelmv && have_ref &&
2091 ref_bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
2092 memcpy(&bsi->rdstat[i][mode_idx], &ref_bsi->rdstat[i][mode_idx],
2093 sizeof(SEG_RDSTAT));
2094 if (num_4x4_blocks_wide > 1)
2095 bsi->rdstat[i + 1][mode_idx].eobs =
2096 ref_bsi->rdstat[i + 1][mode_idx].eobs;
2097 if (num_4x4_blocks_high > 1)
2098 bsi->rdstat[i + 2][mode_idx].eobs =
2099 ref_bsi->rdstat[i + 2][mode_idx].eobs;
2100
2101 if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
2102 mode_selected = this_mode;
2103 best_rd = bsi->rdstat[i][mode_idx].brdcost;
2104 }
2105 continue;
2106 }
2107 }
2108
2109 bsi->rdstat[i][mode_idx].brdcost = encode_inter_mb_segment(
2110 cpi, x, bsi->segment_rd - this_segment_rd, i,
2111 &bsi->rdstat[i][mode_idx].byrate, &bsi->rdstat[i][mode_idx].bdist,
2112 &bsi->rdstat[i][mode_idx].bsse, bsi->rdstat[i][mode_idx].ta,
2113 bsi->rdstat[i][mode_idx].tl, mi_row, mi_col);
2114 if (bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
2115 bsi->rdstat[i][mode_idx].brdcost +=
2116 RDCOST(x->rdmult, x->rddiv, bsi->rdstat[i][mode_idx].brate, 0);
2117 bsi->rdstat[i][mode_idx].brate += bsi->rdstat[i][mode_idx].byrate;
2118 bsi->rdstat[i][mode_idx].eobs = p->eobs[i];
2119 if (num_4x4_blocks_wide > 1)
2120 bsi->rdstat[i + 1][mode_idx].eobs = p->eobs[i + 1];
2121 if (num_4x4_blocks_high > 1)
2122 bsi->rdstat[i + 2][mode_idx].eobs = p->eobs[i + 2];
2123 }
2124
2125 if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
2126 mode_selected = this_mode;
2127 best_rd = bsi->rdstat[i][mode_idx].brdcost;
2128 }
2129 } /*for each 4x4 mode*/
2130
2131 if (best_rd == INT64_MAX) {
2132 int iy, midx;
2133 for (iy = i + 1; iy < 4; ++iy)
2134 for (midx = 0; midx < INTER_MODES; ++midx)
2135 bsi->rdstat[iy][midx].brdcost = INT64_MAX;
2136 bsi->segment_rd = INT64_MAX;
2137 return INT64_MAX;
2138 }
2139
2140 mode_idx = INTER_OFFSET(mode_selected);
2141 memcpy(t_above, bsi->rdstat[i][mode_idx].ta, sizeof(t_above));
2142 memcpy(t_left, bsi->rdstat[i][mode_idx].tl, sizeof(t_left));
2143
2144 set_and_cost_bmi_mvs(cpi, x, xd, i, mode_selected, mode_mv[mode_selected],
2145 frame_mv, seg_mvs[i], bsi->ref_mv, x->nmvjointcost,
2146 x->mvcost);
2147
2148 br += bsi->rdstat[i][mode_idx].brate;
2149 bd += bsi->rdstat[i][mode_idx].bdist;
2150 block_sse += bsi->rdstat[i][mode_idx].bsse;
2151 segmentyrate += bsi->rdstat[i][mode_idx].byrate;
2152 this_segment_rd += bsi->rdstat[i][mode_idx].brdcost;
2153
2154 if (this_segment_rd > bsi->segment_rd) {
2155 int iy, midx;
2156 for (iy = i + 1; iy < 4; ++iy)
2157 for (midx = 0; midx < INTER_MODES; ++midx)
2158 bsi->rdstat[iy][midx].brdcost = INT64_MAX;
2159 bsi->segment_rd = INT64_MAX;
2160 return INT64_MAX;
2161 }
2162 }
2163 } /* for each label */
2164
2165 bsi->r = br;
2166 bsi->d = bd;
2167 bsi->segment_yrate = segmentyrate;
2168 bsi->segment_rd = this_segment_rd;
2169 bsi->sse = block_sse;
2170
2171 // update the coding decisions
2172 for (k = 0; k < 4; ++k) bsi->modes[k] = mi->bmi[k].as_mode;
2173
2174 if (bsi->segment_rd > best_rd) return INT64_MAX;
2175 /* set it to the best */
2176 for (i = 0; i < 4; i++) {
2177 mode_idx = INTER_OFFSET(bsi->modes[i]);
2178 mi->bmi[i].as_mv[0].as_int = bsi->rdstat[i][mode_idx].mvs[0].as_int;
2179 if (has_second_ref(mi))
2180 mi->bmi[i].as_mv[1].as_int = bsi->rdstat[i][mode_idx].mvs[1].as_int;
2181 x->plane[0].eobs[i] = bsi->rdstat[i][mode_idx].eobs;
2182 mi->bmi[i].as_mode = bsi->modes[i];
2183 }
2184
2185 /*
2186 * used to set mbmi->mv.as_int
2187 */
2188 *returntotrate = bsi->r;
2189 *returndistortion = bsi->d;
2190 *returnyrate = bsi->segment_yrate;
2191 *skippable = vp9_is_skippable_in_plane(x, BLOCK_8X8, 0);
2192 *psse = bsi->sse;
2193 mi->mode = bsi->modes[3];
2194
2195 return bsi->segment_rd;
2196 }
2197
estimate_ref_frame_costs(const VP9_COMMON * cm,const MACROBLOCKD * xd,int segment_id,unsigned int * ref_costs_single,unsigned int * ref_costs_comp,vpx_prob * comp_mode_p)2198 static void estimate_ref_frame_costs(const VP9_COMMON *cm,
2199 const MACROBLOCKD *xd, int segment_id,
2200 unsigned int *ref_costs_single,
2201 unsigned int *ref_costs_comp,
2202 vpx_prob *comp_mode_p) {
2203 int seg_ref_active =
2204 segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME);
2205 if (seg_ref_active) {
2206 memset(ref_costs_single, 0, MAX_REF_FRAMES * sizeof(*ref_costs_single));
2207 memset(ref_costs_comp, 0, MAX_REF_FRAMES * sizeof(*ref_costs_comp));
2208 *comp_mode_p = 128;
2209 } else {
2210 vpx_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
2211 vpx_prob comp_inter_p = 128;
2212
2213 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
2214 comp_inter_p = vp9_get_reference_mode_prob(cm, xd);
2215 *comp_mode_p = comp_inter_p;
2216 } else {
2217 *comp_mode_p = 128;
2218 }
2219
2220 ref_costs_single[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
2221
2222 if (cm->reference_mode != COMPOUND_REFERENCE) {
2223 vpx_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
2224 vpx_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
2225 unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
2226
2227 if (cm->reference_mode == REFERENCE_MODE_SELECT)
2228 base_cost += vp9_cost_bit(comp_inter_p, 0);
2229
2230 ref_costs_single[LAST_FRAME] = ref_costs_single[GOLDEN_FRAME] =
2231 ref_costs_single[ALTREF_FRAME] = base_cost;
2232 ref_costs_single[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
2233 ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
2234 ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
2235 ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
2236 ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
2237 } else {
2238 ref_costs_single[LAST_FRAME] = 512;
2239 ref_costs_single[GOLDEN_FRAME] = 512;
2240 ref_costs_single[ALTREF_FRAME] = 512;
2241 }
2242 if (cm->reference_mode != SINGLE_REFERENCE) {
2243 vpx_prob ref_comp_p = vp9_get_pred_prob_comp_ref_p(cm, xd);
2244 unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
2245
2246 if (cm->reference_mode == REFERENCE_MODE_SELECT)
2247 base_cost += vp9_cost_bit(comp_inter_p, 1);
2248
2249 ref_costs_comp[LAST_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 0);
2250 ref_costs_comp[GOLDEN_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 1);
2251 } else {
2252 ref_costs_comp[LAST_FRAME] = 512;
2253 ref_costs_comp[GOLDEN_FRAME] = 512;
2254 }
2255 }
2256 }
2257
store_coding_context(MACROBLOCK * x,PICK_MODE_CONTEXT * ctx,int mode_index,int64_t comp_pred_diff[REFERENCE_MODES],int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS],int skippable)2258 static void store_coding_context(
2259 MACROBLOCK *x, PICK_MODE_CONTEXT *ctx, int mode_index,
2260 int64_t comp_pred_diff[REFERENCE_MODES],
2261 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS], int skippable) {
2262 MACROBLOCKD *const xd = &x->e_mbd;
2263
2264 // Take a snapshot of the coding context so it can be
2265 // restored if we decide to encode this way
2266 ctx->skip = x->skip;
2267 ctx->skippable = skippable;
2268 ctx->best_mode_index = mode_index;
2269 ctx->mic = *xd->mi[0];
2270 ctx->mbmi_ext = *x->mbmi_ext;
2271 ctx->single_pred_diff = (int)comp_pred_diff[SINGLE_REFERENCE];
2272 ctx->comp_pred_diff = (int)comp_pred_diff[COMPOUND_REFERENCE];
2273 ctx->hybrid_pred_diff = (int)comp_pred_diff[REFERENCE_MODE_SELECT];
2274
2275 memcpy(ctx->best_filter_diff, best_filter_diff,
2276 sizeof(*best_filter_diff) * SWITCHABLE_FILTER_CONTEXTS);
2277 }
2278
setup_buffer_inter(VP9_COMP * cpi,MACROBLOCK * x,MV_REFERENCE_FRAME ref_frame,BLOCK_SIZE block_size,int mi_row,int mi_col,int_mv frame_nearest_mv[MAX_REF_FRAMES],int_mv frame_near_mv[MAX_REF_FRAMES],struct buf_2d yv12_mb[4][MAX_MB_PLANE])2279 static void setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
2280 MV_REFERENCE_FRAME ref_frame,
2281 BLOCK_SIZE block_size, int mi_row, int mi_col,
2282 int_mv frame_nearest_mv[MAX_REF_FRAMES],
2283 int_mv frame_near_mv[MAX_REF_FRAMES],
2284 struct buf_2d yv12_mb[4][MAX_MB_PLANE]) {
2285 const VP9_COMMON *cm = &cpi->common;
2286 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
2287 MACROBLOCKD *const xd = &x->e_mbd;
2288 MODE_INFO *const mi = xd->mi[0];
2289 int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
2290 const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
2291 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2292
2293 assert(yv12 != NULL);
2294
2295 // TODO(jkoleszar): Is the UV buffer ever used here? If so, need to make this
2296 // use the UV scaling factors.
2297 vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
2298
2299 // Gets an initial list of candidate vectors from neighbours and orders them
2300 vp9_find_mv_refs(cm, xd, mi, ref_frame, candidates, mi_row, mi_col,
2301 mbmi_ext->mode_context);
2302
2303 // Candidate refinement carried out at encoder and decoder
2304 vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
2305 &frame_nearest_mv[ref_frame],
2306 &frame_near_mv[ref_frame]);
2307
2308 // Further refinement that is encode side only to test the top few candidates
2309 // in full and choose the best as the centre point for subsequent searches.
2310 // The current implementation doesn't support scaling.
2311 if (!vp9_is_scaled(sf) && block_size >= BLOCK_8X8)
2312 vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride, ref_frame,
2313 block_size);
2314 }
2315
single_motion_search(VP9_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,int mi_row,int mi_col,int_mv * tmp_mv,int * rate_mv)2316 static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
2317 int mi_row, int mi_col, int_mv *tmp_mv,
2318 int *rate_mv) {
2319 MACROBLOCKD *xd = &x->e_mbd;
2320 const VP9_COMMON *cm = &cpi->common;
2321 MODE_INFO *mi = xd->mi[0];
2322 struct buf_2d backup_yv12[MAX_MB_PLANE] = { { 0, 0 } };
2323 int bestsme = INT_MAX;
2324 int step_param;
2325 int sadpb = x->sadperbit16;
2326 MV mvp_full;
2327 int ref = mi->ref_frame[0];
2328 MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
2329 const MvLimits tmp_mv_limits = x->mv_limits;
2330 int cost_list[5];
2331
2332 const YV12_BUFFER_CONFIG *scaled_ref_frame =
2333 vp9_get_scaled_ref_frame(cpi, ref);
2334
2335 MV pred_mv[3];
2336 pred_mv[0] = x->mbmi_ext->ref_mvs[ref][0].as_mv;
2337 pred_mv[1] = x->mbmi_ext->ref_mvs[ref][1].as_mv;
2338 pred_mv[2] = x->pred_mv[ref];
2339
2340 if (scaled_ref_frame) {
2341 int i;
2342 // Swap out the reference frame for a version that's been scaled to
2343 // match the resolution of the current frame, allowing the existing
2344 // motion search code to be used without additional modifications.
2345 for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0];
2346
2347 vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
2348 }
2349
2350 // Work out the size of the first step in the mv step search.
2351 // 0 here is maximum length first step. 1 is VPXMAX >> 1 etc.
2352 if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
2353 // Take wtd average of the step_params based on the last frame's
2354 // max mv magnitude and that based on the best ref mvs of the current
2355 // block for the given reference.
2356 step_param =
2357 (vp9_init_search_range(x->max_mv_context[ref]) + cpi->mv_step_param) /
2358 2;
2359 } else {
2360 step_param = cpi->mv_step_param;
2361 }
2362
2363 if (cpi->sf.adaptive_motion_search && bsize < BLOCK_64X64) {
2364 int boffset =
2365 2 * (b_width_log2_lookup[BLOCK_64X64] -
2366 VPXMIN(b_height_log2_lookup[bsize], b_width_log2_lookup[bsize]));
2367 step_param = VPXMAX(step_param, boffset);
2368 }
2369
2370 if (cpi->sf.adaptive_motion_search) {
2371 int bwl = b_width_log2_lookup[bsize];
2372 int bhl = b_height_log2_lookup[bsize];
2373 int tlevel = x->pred_mv_sad[ref] >> (bwl + bhl + 4);
2374
2375 if (tlevel < 5) step_param += 2;
2376
2377 // prev_mv_sad is not setup for dynamically scaled frames.
2378 if (cpi->oxcf.resize_mode != RESIZE_DYNAMIC) {
2379 int i;
2380 for (i = LAST_FRAME; i <= ALTREF_FRAME && cm->show_frame; ++i) {
2381 if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
2382 x->pred_mv[ref].row = 0;
2383 x->pred_mv[ref].col = 0;
2384 tmp_mv->as_int = INVALID_MV;
2385
2386 if (scaled_ref_frame) {
2387 int i;
2388 for (i = 0; i < MAX_MB_PLANE; ++i)
2389 xd->plane[i].pre[0] = backup_yv12[i];
2390 }
2391 return;
2392 }
2393 }
2394 }
2395 }
2396
2397 // Note: MV limits are modified here. Always restore the original values
2398 // after full-pixel motion search.
2399 vp9_set_mv_search_range(&x->mv_limits, &ref_mv);
2400
2401 mvp_full = pred_mv[x->mv_best_ref_index[ref]];
2402
2403 mvp_full.col >>= 3;
2404 mvp_full.row >>= 3;
2405
2406 bestsme = vp9_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb,
2407 cond_cost_list(cpi, cost_list), &ref_mv,
2408 &tmp_mv->as_mv, INT_MAX, 1);
2409
2410 x->mv_limits = tmp_mv_limits;
2411
2412 if (bestsme < INT_MAX) {
2413 uint32_t dis; /* TODO: use dis in distortion calculation later. */
2414 cpi->find_fractional_mv_step(
2415 x, &tmp_mv->as_mv, &ref_mv, cm->allow_high_precision_mv, x->errorperbit,
2416 &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
2417 cpi->sf.mv.subpel_iters_per_step, cond_cost_list(cpi, cost_list),
2418 x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref], NULL, 0, 0);
2419 }
2420 *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv, x->nmvjointcost,
2421 x->mvcost, MV_COST_WEIGHT);
2422
2423 if (cpi->sf.adaptive_motion_search) x->pred_mv[ref] = tmp_mv->as_mv;
2424
2425 if (scaled_ref_frame) {
2426 int i;
2427 for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
2428 }
2429 }
2430
restore_dst_buf(MACROBLOCKD * xd,uint8_t * orig_dst[MAX_MB_PLANE],int orig_dst_stride[MAX_MB_PLANE])2431 static INLINE void restore_dst_buf(MACROBLOCKD *xd,
2432 uint8_t *orig_dst[MAX_MB_PLANE],
2433 int orig_dst_stride[MAX_MB_PLANE]) {
2434 int i;
2435 for (i = 0; i < MAX_MB_PLANE; i++) {
2436 xd->plane[i].dst.buf = orig_dst[i];
2437 xd->plane[i].dst.stride = orig_dst_stride[i];
2438 }
2439 }
2440
2441 // In some situations we want to discount tha pparent cost of a new motion
2442 // vector. Where there is a subtle motion field and especially where there is
2443 // low spatial complexity then it can be hard to cover the cost of a new motion
2444 // vector in a single block, even if that motion vector reduces distortion.
2445 // However, once established that vector may be usable through the nearest and
2446 // near mv modes to reduce distortion in subsequent blocks and also improve
2447 // visual quality.
discount_newmv_test(const VP9_COMP * cpi,int this_mode,int_mv this_mv,int_mv (* mode_mv)[MAX_REF_FRAMES],int ref_frame)2448 static int discount_newmv_test(const VP9_COMP *cpi, int this_mode,
2449 int_mv this_mv,
2450 int_mv (*mode_mv)[MAX_REF_FRAMES],
2451 int ref_frame) {
2452 return (!cpi->rc.is_src_frame_alt_ref && (this_mode == NEWMV) &&
2453 (this_mv.as_int != 0) &&
2454 ((mode_mv[NEARESTMV][ref_frame].as_int == 0) ||
2455 (mode_mv[NEARESTMV][ref_frame].as_int == INVALID_MV)) &&
2456 ((mode_mv[NEARMV][ref_frame].as_int == 0) ||
2457 (mode_mv[NEARMV][ref_frame].as_int == INVALID_MV)));
2458 }
2459
handle_inter_mode(VP9_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,int * rate2,int64_t * distortion,int * skippable,int * rate_y,int * rate_uv,int * disable_skip,int_mv (* mode_mv)[MAX_REF_FRAMES],int mi_row,int mi_col,int_mv single_newmv[MAX_REF_FRAMES],INTERP_FILTER (* single_filter)[MAX_REF_FRAMES],int (* single_skippable)[MAX_REF_FRAMES],int64_t * psse,const int64_t ref_best_rd,int64_t * mask_filter,int64_t filter_cache[])2460 static int64_t handle_inter_mode(
2461 VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize, int *rate2,
2462 int64_t *distortion, int *skippable, int *rate_y, int *rate_uv,
2463 int *disable_skip, int_mv (*mode_mv)[MAX_REF_FRAMES], int mi_row,
2464 int mi_col, int_mv single_newmv[MAX_REF_FRAMES],
2465 INTERP_FILTER (*single_filter)[MAX_REF_FRAMES],
2466 int (*single_skippable)[MAX_REF_FRAMES], int64_t *psse,
2467 const int64_t ref_best_rd, int64_t *mask_filter, int64_t filter_cache[]) {
2468 VP9_COMMON *cm = &cpi->common;
2469 MACROBLOCKD *xd = &x->e_mbd;
2470 MODE_INFO *mi = xd->mi[0];
2471 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2472 const int is_comp_pred = has_second_ref(mi);
2473 const int this_mode = mi->mode;
2474 int_mv *frame_mv = mode_mv[this_mode];
2475 int i;
2476 int refs[2] = { mi->ref_frame[0],
2477 (mi->ref_frame[1] < 0 ? 0 : mi->ref_frame[1]) };
2478 int_mv cur_mv[2];
2479 #if CONFIG_VP9_HIGHBITDEPTH
2480 DECLARE_ALIGNED(16, uint16_t, tmp_buf16[MAX_MB_PLANE * 64 * 64]);
2481 uint8_t *tmp_buf;
2482 #else
2483 DECLARE_ALIGNED(16, uint8_t, tmp_buf[MAX_MB_PLANE * 64 * 64]);
2484 #endif // CONFIG_VP9_HIGHBITDEPTH
2485 int pred_exists = 0;
2486 int intpel_mv;
2487 int64_t rd, tmp_rd, best_rd = INT64_MAX;
2488 int best_needs_copy = 0;
2489 uint8_t *orig_dst[MAX_MB_PLANE];
2490 int orig_dst_stride[MAX_MB_PLANE];
2491 int rs = 0;
2492 INTERP_FILTER best_filter = SWITCHABLE;
2493 uint8_t skip_txfm[MAX_MB_PLANE << 2] = { 0 };
2494 int64_t bsse[MAX_MB_PLANE << 2] = { 0 };
2495
2496 int bsl = mi_width_log2_lookup[bsize];
2497 int pred_filter_search =
2498 cpi->sf.cb_pred_filter_search
2499 ? (((mi_row + mi_col) >> bsl) +
2500 get_chessboard_index(cm->current_video_frame)) &
2501 0x1
2502 : 0;
2503
2504 int skip_txfm_sb = 0;
2505 int64_t skip_sse_sb = INT64_MAX;
2506 int64_t distortion_y = 0, distortion_uv = 0;
2507
2508 #if CONFIG_VP9_HIGHBITDEPTH
2509 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2510 tmp_buf = CONVERT_TO_BYTEPTR(tmp_buf16);
2511 } else {
2512 tmp_buf = (uint8_t *)tmp_buf16;
2513 }
2514 #endif // CONFIG_VP9_HIGHBITDEPTH
2515
2516 if (pred_filter_search) {
2517 INTERP_FILTER af = SWITCHABLE, lf = SWITCHABLE;
2518 if (xd->above_mi && is_inter_block(xd->above_mi))
2519 af = xd->above_mi->interp_filter;
2520 if (xd->left_mi && is_inter_block(xd->left_mi))
2521 lf = xd->left_mi->interp_filter;
2522
2523 if ((this_mode != NEWMV) || (af == lf)) best_filter = af;
2524 }
2525
2526 if (is_comp_pred) {
2527 if (frame_mv[refs[0]].as_int == INVALID_MV ||
2528 frame_mv[refs[1]].as_int == INVALID_MV)
2529 return INT64_MAX;
2530
2531 if (cpi->sf.adaptive_mode_search) {
2532 if (single_filter[this_mode][refs[0]] ==
2533 single_filter[this_mode][refs[1]])
2534 best_filter = single_filter[this_mode][refs[0]];
2535 }
2536 }
2537
2538 if (this_mode == NEWMV) {
2539 int rate_mv;
2540 if (is_comp_pred) {
2541 // Initialize mv using single prediction mode result.
2542 frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int;
2543 frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int;
2544
2545 if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
2546 joint_motion_search(cpi, x, bsize, frame_mv, mi_row, mi_col,
2547 single_newmv, &rate_mv);
2548 } else {
2549 rate_mv = vp9_mv_bit_cost(&frame_mv[refs[0]].as_mv,
2550 &x->mbmi_ext->ref_mvs[refs[0]][0].as_mv,
2551 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2552 rate_mv += vp9_mv_bit_cost(&frame_mv[refs[1]].as_mv,
2553 &x->mbmi_ext->ref_mvs[refs[1]][0].as_mv,
2554 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2555 }
2556 *rate2 += rate_mv;
2557 } else {
2558 int_mv tmp_mv;
2559 single_motion_search(cpi, x, bsize, mi_row, mi_col, &tmp_mv, &rate_mv);
2560 if (tmp_mv.as_int == INVALID_MV) return INT64_MAX;
2561
2562 frame_mv[refs[0]].as_int = xd->mi[0]->bmi[0].as_mv[0].as_int =
2563 tmp_mv.as_int;
2564 single_newmv[refs[0]].as_int = tmp_mv.as_int;
2565
2566 // Estimate the rate implications of a new mv but discount this
2567 // under certain circumstances where we want to help initiate a weak
2568 // motion field, where the distortion gain for a single block may not
2569 // be enough to overcome the cost of a new mv.
2570 if (discount_newmv_test(cpi, this_mode, tmp_mv, mode_mv, refs[0])) {
2571 *rate2 += VPXMAX((rate_mv / NEW_MV_DISCOUNT_FACTOR), 1);
2572 } else {
2573 *rate2 += rate_mv;
2574 }
2575 }
2576 }
2577
2578 for (i = 0; i < is_comp_pred + 1; ++i) {
2579 cur_mv[i] = frame_mv[refs[i]];
2580 // Clip "next_nearest" so that it does not extend to far out of image
2581 if (this_mode != NEWMV) clamp_mv2(&cur_mv[i].as_mv, xd);
2582
2583 if (mv_check_bounds(&x->mv_limits, &cur_mv[i].as_mv)) return INT64_MAX;
2584 mi->mv[i].as_int = cur_mv[i].as_int;
2585 }
2586
2587 // do first prediction into the destination buffer. Do the next
2588 // prediction into a temporary buffer. Then keep track of which one
2589 // of these currently holds the best predictor, and use the other
2590 // one for future predictions. In the end, copy from tmp_buf to
2591 // dst if necessary.
2592 for (i = 0; i < MAX_MB_PLANE; i++) {
2593 orig_dst[i] = xd->plane[i].dst.buf;
2594 orig_dst_stride[i] = xd->plane[i].dst.stride;
2595 }
2596
2597 // We don't include the cost of the second reference here, because there
2598 // are only two options: Last/ARF or Golden/ARF; The second one is always
2599 // known, which is ARF.
2600 //
2601 // Under some circumstances we discount the cost of new mv mode to encourage
2602 // initiation of a motion field.
2603 if (discount_newmv_test(cpi, this_mode, frame_mv[refs[0]], mode_mv,
2604 refs[0])) {
2605 *rate2 +=
2606 VPXMIN(cost_mv_ref(cpi, this_mode, mbmi_ext->mode_context[refs[0]]),
2607 cost_mv_ref(cpi, NEARESTMV, mbmi_ext->mode_context[refs[0]]));
2608 } else {
2609 *rate2 += cost_mv_ref(cpi, this_mode, mbmi_ext->mode_context[refs[0]]);
2610 }
2611
2612 if (RDCOST(x->rdmult, x->rddiv, *rate2, 0) > ref_best_rd &&
2613 mi->mode != NEARESTMV)
2614 return INT64_MAX;
2615
2616 pred_exists = 0;
2617 // Are all MVs integer pel for Y and UV
2618 intpel_mv = !mv_has_subpel(&mi->mv[0].as_mv);
2619 if (is_comp_pred) intpel_mv &= !mv_has_subpel(&mi->mv[1].as_mv);
2620
2621 // Search for best switchable filter by checking the variance of
2622 // pred error irrespective of whether the filter will be used
2623 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i) filter_cache[i] = INT64_MAX;
2624
2625 if (cm->interp_filter != BILINEAR) {
2626 if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) {
2627 best_filter = EIGHTTAP;
2628 } else if (best_filter == SWITCHABLE) {
2629 int newbest;
2630 int tmp_rate_sum = 0;
2631 int64_t tmp_dist_sum = 0;
2632
2633 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
2634 int j;
2635 int64_t rs_rd;
2636 int tmp_skip_sb = 0;
2637 int64_t tmp_skip_sse = INT64_MAX;
2638
2639 mi->interp_filter = i;
2640 rs = vp9_get_switchable_rate(cpi, xd);
2641 rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
2642
2643 if (i > 0 && intpel_mv) {
2644 rd = RDCOST(x->rdmult, x->rddiv, tmp_rate_sum, tmp_dist_sum);
2645 filter_cache[i] = rd;
2646 filter_cache[SWITCHABLE_FILTERS] =
2647 VPXMIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2648 if (cm->interp_filter == SWITCHABLE) rd += rs_rd;
2649 *mask_filter = VPXMAX(*mask_filter, rd);
2650 } else {
2651 int rate_sum = 0;
2652 int64_t dist_sum = 0;
2653 if (i > 0 && cpi->sf.adaptive_interp_filter_search &&
2654 (cpi->sf.interp_filter_search_mask & (1 << i))) {
2655 rate_sum = INT_MAX;
2656 dist_sum = INT64_MAX;
2657 continue;
2658 }
2659
2660 if ((cm->interp_filter == SWITCHABLE && (!i || best_needs_copy)) ||
2661 (cm->interp_filter != SWITCHABLE &&
2662 (cm->interp_filter == mi->interp_filter ||
2663 (i == 0 && intpel_mv)))) {
2664 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2665 } else {
2666 for (j = 0; j < MAX_MB_PLANE; j++) {
2667 xd->plane[j].dst.buf = tmp_buf + j * 64 * 64;
2668 xd->plane[j].dst.stride = 64;
2669 }
2670 }
2671 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2672 model_rd_for_sb(cpi, bsize, x, xd, &rate_sum, &dist_sum, &tmp_skip_sb,
2673 &tmp_skip_sse);
2674
2675 rd = RDCOST(x->rdmult, x->rddiv, rate_sum, dist_sum);
2676 filter_cache[i] = rd;
2677 filter_cache[SWITCHABLE_FILTERS] =
2678 VPXMIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2679 if (cm->interp_filter == SWITCHABLE) rd += rs_rd;
2680 *mask_filter = VPXMAX(*mask_filter, rd);
2681
2682 if (i == 0 && intpel_mv) {
2683 tmp_rate_sum = rate_sum;
2684 tmp_dist_sum = dist_sum;
2685 }
2686 }
2687
2688 if (i == 0 && cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2689 if (rd / 2 > ref_best_rd) {
2690 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2691 return INT64_MAX;
2692 }
2693 }
2694 newbest = i == 0 || rd < best_rd;
2695
2696 if (newbest) {
2697 best_rd = rd;
2698 best_filter = mi->interp_filter;
2699 if (cm->interp_filter == SWITCHABLE && i && !intpel_mv)
2700 best_needs_copy = !best_needs_copy;
2701 }
2702
2703 if ((cm->interp_filter == SWITCHABLE && newbest) ||
2704 (cm->interp_filter != SWITCHABLE &&
2705 cm->interp_filter == mi->interp_filter)) {
2706 pred_exists = 1;
2707 tmp_rd = best_rd;
2708
2709 skip_txfm_sb = tmp_skip_sb;
2710 skip_sse_sb = tmp_skip_sse;
2711 memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2712 memcpy(bsse, x->bsse, sizeof(bsse));
2713 }
2714 }
2715 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2716 }
2717 }
2718 // Set the appropriate filter
2719 mi->interp_filter =
2720 cm->interp_filter != SWITCHABLE ? cm->interp_filter : best_filter;
2721 rs = cm->interp_filter == SWITCHABLE ? vp9_get_switchable_rate(cpi, xd) : 0;
2722
2723 if (pred_exists) {
2724 if (best_needs_copy) {
2725 // again temporarily set the buffers to local memory to prevent a memcpy
2726 for (i = 0; i < MAX_MB_PLANE; i++) {
2727 xd->plane[i].dst.buf = tmp_buf + i * 64 * 64;
2728 xd->plane[i].dst.stride = 64;
2729 }
2730 }
2731 rd = tmp_rd + RDCOST(x->rdmult, x->rddiv, rs, 0);
2732 } else {
2733 int tmp_rate;
2734 int64_t tmp_dist;
2735 // Handles the special case when a filter that is not in the
2736 // switchable list (ex. bilinear) is indicated at the frame level, or
2737 // skip condition holds.
2738 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2739 model_rd_for_sb(cpi, bsize, x, xd, &tmp_rate, &tmp_dist, &skip_txfm_sb,
2740 &skip_sse_sb);
2741 rd = RDCOST(x->rdmult, x->rddiv, rs + tmp_rate, tmp_dist);
2742 memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
2743 memcpy(bsse, x->bsse, sizeof(bsse));
2744 }
2745
2746 if (!is_comp_pred) single_filter[this_mode][refs[0]] = mi->interp_filter;
2747
2748 if (cpi->sf.adaptive_mode_search)
2749 if (is_comp_pred)
2750 if (single_skippable[this_mode][refs[0]] &&
2751 single_skippable[this_mode][refs[1]])
2752 memset(skip_txfm, SKIP_TXFM_AC_DC, sizeof(skip_txfm));
2753
2754 if (cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2755 // if current pred_error modeled rd is substantially more than the best
2756 // so far, do not bother doing full rd
2757 if (rd / 2 > ref_best_rd) {
2758 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2759 return INT64_MAX;
2760 }
2761 }
2762
2763 if (cm->interp_filter == SWITCHABLE) *rate2 += rs;
2764
2765 memcpy(x->skip_txfm, skip_txfm, sizeof(skip_txfm));
2766 memcpy(x->bsse, bsse, sizeof(bsse));
2767
2768 if (!skip_txfm_sb) {
2769 int skippable_y, skippable_uv;
2770 int64_t sseuv = INT64_MAX;
2771 int64_t rdcosty = INT64_MAX;
2772
2773 // Y cost and distortion
2774 vp9_subtract_plane(x, bsize, 0);
2775 super_block_yrd(cpi, x, rate_y, &distortion_y, &skippable_y, psse, bsize,
2776 ref_best_rd);
2777
2778 if (*rate_y == INT_MAX) {
2779 *rate2 = INT_MAX;
2780 *distortion = INT64_MAX;
2781 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2782 return INT64_MAX;
2783 }
2784
2785 *rate2 += *rate_y;
2786 *distortion += distortion_y;
2787
2788 rdcosty = RDCOST(x->rdmult, x->rddiv, *rate2, *distortion);
2789 rdcosty = VPXMIN(rdcosty, RDCOST(x->rdmult, x->rddiv, 0, *psse));
2790
2791 if (!super_block_uvrd(cpi, x, rate_uv, &distortion_uv, &skippable_uv,
2792 &sseuv, bsize, ref_best_rd - rdcosty)) {
2793 *rate2 = INT_MAX;
2794 *distortion = INT64_MAX;
2795 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2796 return INT64_MAX;
2797 }
2798
2799 *psse += sseuv;
2800 *rate2 += *rate_uv;
2801 *distortion += distortion_uv;
2802 *skippable = skippable_y && skippable_uv;
2803 } else {
2804 x->skip = 1;
2805 *disable_skip = 1;
2806
2807 // The cost of skip bit needs to be added.
2808 *rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2809
2810 *distortion = skip_sse_sb;
2811 }
2812
2813 if (!is_comp_pred) single_skippable[this_mode][refs[0]] = *skippable;
2814
2815 restore_dst_buf(xd, orig_dst, orig_dst_stride);
2816 return 0; // The rate-distortion cost will be re-calculated by caller.
2817 }
2818
vp9_rd_pick_intra_mode_sb(VP9_COMP * cpi,MACROBLOCK * x,RD_COST * rd_cost,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx,int64_t best_rd)2819 void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
2820 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
2821 int64_t best_rd) {
2822 VP9_COMMON *const cm = &cpi->common;
2823 MACROBLOCKD *const xd = &x->e_mbd;
2824 struct macroblockd_plane *const pd = xd->plane;
2825 int rate_y = 0, rate_uv = 0, rate_y_tokenonly = 0, rate_uv_tokenonly = 0;
2826 int y_skip = 0, uv_skip = 0;
2827 int64_t dist_y = 0, dist_uv = 0;
2828 TX_SIZE max_uv_tx_size;
2829 x->skip_encode = 0;
2830 ctx->skip = 0;
2831 xd->mi[0]->ref_frame[0] = INTRA_FRAME;
2832 xd->mi[0]->ref_frame[1] = NONE;
2833 // Initialize interp_filter here so we do not have to check for inter block
2834 // modes in get_pred_context_switchable_interp()
2835 xd->mi[0]->interp_filter = SWITCHABLE_FILTERS;
2836
2837 if (bsize >= BLOCK_8X8) {
2838 if (rd_pick_intra_sby_mode(cpi, x, &rate_y, &rate_y_tokenonly, &dist_y,
2839 &y_skip, bsize, best_rd) >= best_rd) {
2840 rd_cost->rate = INT_MAX;
2841 return;
2842 }
2843 } else {
2844 y_skip = 0;
2845 if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate_y, &rate_y_tokenonly,
2846 &dist_y, best_rd) >= best_rd) {
2847 rd_cost->rate = INT_MAX;
2848 return;
2849 }
2850 }
2851 max_uv_tx_size = uv_txsize_lookup[bsize][xd->mi[0]->tx_size]
2852 [pd[1].subsampling_x][pd[1].subsampling_y];
2853 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly, &dist_uv,
2854 &uv_skip, VPXMAX(BLOCK_8X8, bsize), max_uv_tx_size);
2855
2856 if (y_skip && uv_skip) {
2857 rd_cost->rate = rate_y + rate_uv - rate_y_tokenonly - rate_uv_tokenonly +
2858 vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2859 rd_cost->dist = dist_y + dist_uv;
2860 } else {
2861 rd_cost->rate =
2862 rate_y + rate_uv + vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
2863 rd_cost->dist = dist_y + dist_uv;
2864 }
2865
2866 ctx->mic = *xd->mi[0];
2867 ctx->mbmi_ext = *x->mbmi_ext;
2868 rd_cost->rdcost = RDCOST(x->rdmult, x->rddiv, rd_cost->rate, rd_cost->dist);
2869 }
2870
2871 // This function is designed to apply a bias or adjustment to an rd value based
2872 // on the relative variance of the source and reconstruction.
2873 #define LOW_VAR_THRESH 16
2874 #define VLOW_ADJ_MAX 25
2875 #define VHIGH_ADJ_MAX 8
rd_variance_adjustment(VP9_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,int64_t * this_rd,MV_REFERENCE_FRAME ref_frame,unsigned int source_variance)2876 static void rd_variance_adjustment(VP9_COMP *cpi, MACROBLOCK *x,
2877 BLOCK_SIZE bsize, int64_t *this_rd,
2878 MV_REFERENCE_FRAME ref_frame,
2879 unsigned int source_variance) {
2880 MACROBLOCKD *const xd = &x->e_mbd;
2881 unsigned int recon_variance;
2882 unsigned int absvar_diff = 0;
2883 int64_t var_error = 0;
2884 int64_t var_factor = 0;
2885
2886 if (*this_rd == INT64_MAX) return;
2887
2888 #if CONFIG_VP9_HIGHBITDEPTH
2889 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2890 recon_variance = vp9_high_get_sby_perpixel_variance(cpi, &xd->plane[0].dst,
2891 bsize, xd->bd);
2892 } else {
2893 recon_variance =
2894 vp9_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize);
2895 }
2896 #else
2897 recon_variance = vp9_get_sby_perpixel_variance(cpi, &xd->plane[0].dst, bsize);
2898 #endif // CONFIG_VP9_HIGHBITDEPTH
2899
2900 if ((source_variance + recon_variance) > LOW_VAR_THRESH) {
2901 absvar_diff = (source_variance > recon_variance)
2902 ? (source_variance - recon_variance)
2903 : (recon_variance - source_variance);
2904
2905 var_error = ((int64_t)200 * source_variance * recon_variance) /
2906 (((int64_t)source_variance * source_variance) +
2907 ((int64_t)recon_variance * recon_variance));
2908 var_error = 100 - var_error;
2909 }
2910
2911 // Source variance above a threshold and ref frame is intra.
2912 // This case is targeted mainly at discouraging intra modes that give rise
2913 // to a predictor with a low spatial complexity compared to the source.
2914 if ((source_variance > LOW_VAR_THRESH) && (ref_frame == INTRA_FRAME) &&
2915 (source_variance > recon_variance)) {
2916 var_factor = VPXMIN(absvar_diff, VPXMIN(VLOW_ADJ_MAX, var_error));
2917 // A second possible case of interest is where the source variance
2918 // is very low and we wish to discourage false texture or motion trails.
2919 } else if ((source_variance < (LOW_VAR_THRESH >> 1)) &&
2920 (recon_variance > source_variance)) {
2921 var_factor = VPXMIN(absvar_diff, VPXMIN(VHIGH_ADJ_MAX, var_error));
2922 }
2923 *this_rd += (*this_rd * var_factor) / 100;
2924 }
2925
2926 // Do we have an internal image edge (e.g. formatting bars).
vp9_internal_image_edge(VP9_COMP * cpi)2927 int vp9_internal_image_edge(VP9_COMP *cpi) {
2928 return (cpi->oxcf.pass == 2) &&
2929 ((cpi->twopass.this_frame_stats.inactive_zone_rows > 0) ||
2930 (cpi->twopass.this_frame_stats.inactive_zone_cols > 0));
2931 }
2932
2933 // Checks to see if a super block is on a horizontal image edge.
2934 // In most cases this is the "real" edge unless there are formatting
2935 // bars embedded in the stream.
vp9_active_h_edge(VP9_COMP * cpi,int mi_row,int mi_step)2936 int vp9_active_h_edge(VP9_COMP *cpi, int mi_row, int mi_step) {
2937 int top_edge = 0;
2938 int bottom_edge = cpi->common.mi_rows;
2939 int is_active_h_edge = 0;
2940
2941 // For two pass account for any formatting bars detected.
2942 if (cpi->oxcf.pass == 2) {
2943 TWO_PASS *twopass = &cpi->twopass;
2944
2945 // The inactive region is specified in MBs not mi units.
2946 // The image edge is in the following MB row.
2947 top_edge += (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
2948
2949 bottom_edge -= (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
2950 bottom_edge = VPXMAX(top_edge, bottom_edge);
2951 }
2952
2953 if (((top_edge >= mi_row) && (top_edge < (mi_row + mi_step))) ||
2954 ((bottom_edge >= mi_row) && (bottom_edge < (mi_row + mi_step)))) {
2955 is_active_h_edge = 1;
2956 }
2957 return is_active_h_edge;
2958 }
2959
2960 // Checks to see if a super block is on a vertical image edge.
2961 // In most cases this is the "real" edge unless there are formatting
2962 // bars embedded in the stream.
vp9_active_v_edge(VP9_COMP * cpi,int mi_col,int mi_step)2963 int vp9_active_v_edge(VP9_COMP *cpi, int mi_col, int mi_step) {
2964 int left_edge = 0;
2965 int right_edge = cpi->common.mi_cols;
2966 int is_active_v_edge = 0;
2967
2968 // For two pass account for any formatting bars detected.
2969 if (cpi->oxcf.pass == 2) {
2970 TWO_PASS *twopass = &cpi->twopass;
2971
2972 // The inactive region is specified in MBs not mi units.
2973 // The image edge is in the following MB row.
2974 left_edge += (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
2975
2976 right_edge -= (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
2977 right_edge = VPXMAX(left_edge, right_edge);
2978 }
2979
2980 if (((left_edge >= mi_col) && (left_edge < (mi_col + mi_step))) ||
2981 ((right_edge >= mi_col) && (right_edge < (mi_col + mi_step)))) {
2982 is_active_v_edge = 1;
2983 }
2984 return is_active_v_edge;
2985 }
2986
2987 // Checks to see if a super block is at the edge of the active image.
2988 // In most cases this is the "real" edge unless there are formatting
2989 // bars embedded in the stream.
vp9_active_edge_sb(VP9_COMP * cpi,int mi_row,int mi_col)2990 int vp9_active_edge_sb(VP9_COMP *cpi, int mi_row, int mi_col) {
2991 return vp9_active_h_edge(cpi, mi_row, MI_BLOCK_SIZE) ||
2992 vp9_active_v_edge(cpi, mi_col, MI_BLOCK_SIZE);
2993 }
2994
vp9_rd_pick_inter_mode_sb(VP9_COMP * cpi,TileDataEnc * tile_data,MACROBLOCK * x,int mi_row,int mi_col,RD_COST * rd_cost,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx,int64_t best_rd_so_far)2995 void vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, TileDataEnc *tile_data,
2996 MACROBLOCK *x, int mi_row, int mi_col,
2997 RD_COST *rd_cost, BLOCK_SIZE bsize,
2998 PICK_MODE_CONTEXT *ctx, int64_t best_rd_so_far) {
2999 VP9_COMMON *const cm = &cpi->common;
3000 TileInfo *const tile_info = &tile_data->tile_info;
3001 RD_OPT *const rd_opt = &cpi->rd;
3002 SPEED_FEATURES *const sf = &cpi->sf;
3003 MACROBLOCKD *const xd = &x->e_mbd;
3004 MODE_INFO *const mi = xd->mi[0];
3005 MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
3006 const struct segmentation *const seg = &cm->seg;
3007 PREDICTION_MODE this_mode;
3008 MV_REFERENCE_FRAME ref_frame, second_ref_frame;
3009 unsigned char segment_id = mi->segment_id;
3010 int comp_pred, i, k;
3011 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
3012 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
3013 int_mv single_newmv[MAX_REF_FRAMES] = { { 0 } };
3014 INTERP_FILTER single_inter_filter[MB_MODE_COUNT][MAX_REF_FRAMES];
3015 int single_skippable[MB_MODE_COUNT][MAX_REF_FRAMES];
3016 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
3017 VP9_ALT_FLAG };
3018 int64_t best_rd = best_rd_so_far;
3019 int64_t best_pred_diff[REFERENCE_MODES];
3020 int64_t best_pred_rd[REFERENCE_MODES];
3021 int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
3022 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3023 MODE_INFO best_mbmode;
3024 int best_mode_skippable = 0;
3025 int midx, best_mode_index = -1;
3026 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3027 vpx_prob comp_mode_p;
3028 int64_t best_intra_rd = INT64_MAX;
3029 unsigned int best_pred_sse = UINT_MAX;
3030 PREDICTION_MODE best_intra_mode = DC_PRED;
3031 int rate_uv_intra[TX_SIZES], rate_uv_tokenonly[TX_SIZES];
3032 int64_t dist_uv[TX_SIZES];
3033 int skip_uv[TX_SIZES];
3034 PREDICTION_MODE mode_uv[TX_SIZES];
3035 const int intra_cost_penalty = vp9_get_intra_cost_penalty(
3036 cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
3037 int best_skip2 = 0;
3038 uint8_t ref_frame_skip_mask[2] = { 0 };
3039 uint16_t mode_skip_mask[MAX_REF_FRAMES] = { 0 };
3040 int mode_skip_start = sf->mode_skip_start + 1;
3041 const int *const rd_threshes = rd_opt->threshes[segment_id][bsize];
3042 const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
3043 int64_t mode_threshold[MAX_MODES];
3044 int *mode_map = tile_data->mode_map[bsize];
3045 const int mode_search_skip_flags = sf->mode_search_skip_flags;
3046 int64_t mask_filter = 0;
3047 int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
3048
3049 vp9_zero(best_mbmode);
3050
3051 x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3052
3053 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i) filter_cache[i] = INT64_MAX;
3054
3055 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3056 &comp_mode_p);
3057
3058 for (i = 0; i < REFERENCE_MODES; ++i) best_pred_rd[i] = INT64_MAX;
3059 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3060 best_filter_rd[i] = INT64_MAX;
3061 for (i = 0; i < TX_SIZES; i++) rate_uv_intra[i] = INT_MAX;
3062 for (i = 0; i < MAX_REF_FRAMES; ++i) x->pred_sse[i] = INT_MAX;
3063 for (i = 0; i < MB_MODE_COUNT; ++i) {
3064 for (k = 0; k < MAX_REF_FRAMES; ++k) {
3065 single_inter_filter[i][k] = SWITCHABLE;
3066 single_skippable[i][k] = 0;
3067 }
3068 }
3069
3070 rd_cost->rate = INT_MAX;
3071
3072 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3073 x->pred_mv_sad[ref_frame] = INT_MAX;
3074 if (cpi->ref_frame_flags & flag_list[ref_frame]) {
3075 assert(get_ref_frame_buffer(cpi, ref_frame) != NULL);
3076 setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
3077 frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
3078 }
3079 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
3080 frame_mv[ZEROMV][ref_frame].as_int = 0;
3081 }
3082
3083 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3084 if (!(cpi->ref_frame_flags & flag_list[ref_frame])) {
3085 // Skip checking missing references in both single and compound reference
3086 // modes. Note that a mode will be skipped if both reference frames
3087 // are masked out.
3088 ref_frame_skip_mask[0] |= (1 << ref_frame);
3089 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3090 } else if (sf->reference_masking) {
3091 for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
3092 // Skip fixed mv modes for poor references
3093 if ((x->pred_mv_sad[ref_frame] >> 2) > x->pred_mv_sad[i]) {
3094 mode_skip_mask[ref_frame] |= INTER_NEAREST_NEAR_ZERO;
3095 break;
3096 }
3097 }
3098 }
3099 // If the segment reference frame feature is enabled....
3100 // then do nothing if the current ref frame is not allowed..
3101 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
3102 get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
3103 ref_frame_skip_mask[0] |= (1 << ref_frame);
3104 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3105 }
3106 }
3107
3108 // Disable this drop out case if the ref frame
3109 // segment level feature is enabled for this segment. This is to
3110 // prevent the possibility that we end up unable to pick any mode.
3111 if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
3112 // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
3113 // unless ARNR filtering is enabled in which case we want
3114 // an unfiltered alternative. We allow near/nearest as well
3115 // because they may result in zero-zero MVs but be cheaper.
3116 if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) {
3117 ref_frame_skip_mask[0] = (1 << LAST_FRAME) | (1 << GOLDEN_FRAME);
3118 ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
3119 mode_skip_mask[ALTREF_FRAME] = ~INTER_NEAREST_NEAR_ZERO;
3120 if (frame_mv[NEARMV][ALTREF_FRAME].as_int != 0)
3121 mode_skip_mask[ALTREF_FRAME] |= (1 << NEARMV);
3122 if (frame_mv[NEARESTMV][ALTREF_FRAME].as_int != 0)
3123 mode_skip_mask[ALTREF_FRAME] |= (1 << NEARESTMV);
3124 }
3125 }
3126
3127 if (cpi->rc.is_src_frame_alt_ref) {
3128 if (sf->alt_ref_search_fp) {
3129 mode_skip_mask[ALTREF_FRAME] = 0;
3130 ref_frame_skip_mask[0] = ~(1 << ALTREF_FRAME);
3131 ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
3132 }
3133 }
3134
3135 if (sf->alt_ref_search_fp)
3136 if (!cm->show_frame && x->pred_mv_sad[GOLDEN_FRAME] < INT_MAX)
3137 if (x->pred_mv_sad[ALTREF_FRAME] > (x->pred_mv_sad[GOLDEN_FRAME] << 1))
3138 mode_skip_mask[ALTREF_FRAME] |= INTER_ALL;
3139
3140 if (sf->adaptive_mode_search) {
3141 if (cm->show_frame && !cpi->rc.is_src_frame_alt_ref &&
3142 cpi->rc.frames_since_golden >= 3)
3143 if (x->pred_mv_sad[GOLDEN_FRAME] > (x->pred_mv_sad[LAST_FRAME] << 1))
3144 mode_skip_mask[GOLDEN_FRAME] |= INTER_ALL;
3145 }
3146
3147 if (bsize > sf->max_intra_bsize) {
3148 ref_frame_skip_mask[0] |= (1 << INTRA_FRAME);
3149 ref_frame_skip_mask[1] |= (1 << INTRA_FRAME);
3150 }
3151
3152 mode_skip_mask[INTRA_FRAME] |=
3153 ~(sf->intra_y_mode_mask[max_txsize_lookup[bsize]]);
3154
3155 for (i = 0; i <= LAST_NEW_MV_INDEX; ++i) mode_threshold[i] = 0;
3156 for (i = LAST_NEW_MV_INDEX + 1; i < MAX_MODES; ++i)
3157 mode_threshold[i] = ((int64_t)rd_threshes[i] * rd_thresh_freq_fact[i]) >> 5;
3158
3159 midx = sf->schedule_mode_search ? mode_skip_start : 0;
3160 while (midx > 4) {
3161 uint8_t end_pos = 0;
3162 for (i = 5; i < midx; ++i) {
3163 if (mode_threshold[mode_map[i - 1]] > mode_threshold[mode_map[i]]) {
3164 uint8_t tmp = mode_map[i];
3165 mode_map[i] = mode_map[i - 1];
3166 mode_map[i - 1] = tmp;
3167 end_pos = i;
3168 }
3169 }
3170 midx = end_pos;
3171 }
3172
3173 for (midx = 0; midx < MAX_MODES; ++midx) {
3174 int mode_index = mode_map[midx];
3175 int mode_excluded = 0;
3176 int64_t this_rd = INT64_MAX;
3177 int disable_skip = 0;
3178 int compmode_cost = 0;
3179 int rate2 = 0, rate_y = 0, rate_uv = 0;
3180 int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
3181 int skippable = 0;
3182 int this_skip2 = 0;
3183 int64_t total_sse = INT64_MAX;
3184 int early_term = 0;
3185
3186 this_mode = vp9_mode_order[mode_index].mode;
3187 ref_frame = vp9_mode_order[mode_index].ref_frame[0];
3188 second_ref_frame = vp9_mode_order[mode_index].ref_frame[1];
3189
3190 // Look at the reference frame of the best mode so far and set the
3191 // skip mask to look at a subset of the remaining modes.
3192 if (midx == mode_skip_start && best_mode_index >= 0) {
3193 switch (best_mbmode.ref_frame[0]) {
3194 case INTRA_FRAME: break;
3195 case LAST_FRAME:
3196 ref_frame_skip_mask[0] |= LAST_FRAME_MODE_MASK;
3197 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3198 break;
3199 case GOLDEN_FRAME:
3200 ref_frame_skip_mask[0] |= GOLDEN_FRAME_MODE_MASK;
3201 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3202 break;
3203 case ALTREF_FRAME: ref_frame_skip_mask[0] |= ALT_REF_MODE_MASK; break;
3204 case NONE:
3205 case MAX_REF_FRAMES: assert(0 && "Invalid Reference frame"); break;
3206 }
3207 }
3208
3209 if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
3210 (ref_frame_skip_mask[1] & (1 << VPXMAX(0, second_ref_frame))))
3211 continue;
3212
3213 if (mode_skip_mask[ref_frame] & (1 << this_mode)) continue;
3214
3215 // Test best rd so far against threshold for trying this mode.
3216 if (best_mode_skippable && sf->schedule_mode_search)
3217 mode_threshold[mode_index] <<= 1;
3218
3219 if (best_rd < mode_threshold[mode_index]) continue;
3220
3221 if (sf->motion_field_mode_search) {
3222 const int mi_width = VPXMIN(num_8x8_blocks_wide_lookup[bsize],
3223 tile_info->mi_col_end - mi_col);
3224 const int mi_height = VPXMIN(num_8x8_blocks_high_lookup[bsize],
3225 tile_info->mi_row_end - mi_row);
3226 const int bsl = mi_width_log2_lookup[bsize];
3227 int cb_partition_search_ctrl =
3228 (((mi_row + mi_col) >> bsl) +
3229 get_chessboard_index(cm->current_video_frame)) &
3230 0x1;
3231 MODE_INFO *ref_mi;
3232 int const_motion = 1;
3233 int skip_ref_frame = !cb_partition_search_ctrl;
3234 MV_REFERENCE_FRAME rf = NONE;
3235 int_mv ref_mv;
3236 ref_mv.as_int = INVALID_MV;
3237
3238 if ((mi_row - 1) >= tile_info->mi_row_start) {
3239 ref_mv = xd->mi[-xd->mi_stride]->mv[0];
3240 rf = xd->mi[-xd->mi_stride]->ref_frame[0];
3241 for (i = 0; i < mi_width; ++i) {
3242 ref_mi = xd->mi[-xd->mi_stride + i];
3243 const_motion &= (ref_mv.as_int == ref_mi->mv[0].as_int) &&
3244 (ref_frame == ref_mi->ref_frame[0]);
3245 skip_ref_frame &= (rf == ref_mi->ref_frame[0]);
3246 }
3247 }
3248
3249 if ((mi_col - 1) >= tile_info->mi_col_start) {
3250 if (ref_mv.as_int == INVALID_MV) ref_mv = xd->mi[-1]->mv[0];
3251 if (rf == NONE) rf = xd->mi[-1]->ref_frame[0];
3252 for (i = 0; i < mi_height; ++i) {
3253 ref_mi = xd->mi[i * xd->mi_stride - 1];
3254 const_motion &= (ref_mv.as_int == ref_mi->mv[0].as_int) &&
3255 (ref_frame == ref_mi->ref_frame[0]);
3256 skip_ref_frame &= (rf == ref_mi->ref_frame[0]);
3257 }
3258 }
3259
3260 if (skip_ref_frame && this_mode != NEARESTMV && this_mode != NEWMV)
3261 if (rf > INTRA_FRAME)
3262 if (ref_frame != rf) continue;
3263
3264 if (const_motion)
3265 if (this_mode == NEARMV || this_mode == ZEROMV) continue;
3266 }
3267
3268 comp_pred = second_ref_frame > INTRA_FRAME;
3269 if (comp_pred) {
3270 if (!cpi->allow_comp_inter_inter) continue;
3271
3272 // Skip compound inter modes if ARF is not available.
3273 if (!(cpi->ref_frame_flags & flag_list[second_ref_frame])) continue;
3274
3275 // Do not allow compound prediction if the segment level reference frame
3276 // feature is in use as in this case there can only be one reference.
3277 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) continue;
3278
3279 if ((mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
3280 best_mode_index >= 0 && best_mbmode.ref_frame[0] == INTRA_FRAME)
3281 continue;
3282
3283 mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
3284 } else {
3285 if (ref_frame != INTRA_FRAME)
3286 mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
3287 }
3288
3289 if (ref_frame == INTRA_FRAME) {
3290 if (sf->adaptive_mode_search)
3291 if ((x->source_variance << num_pels_log2_lookup[bsize]) > best_pred_sse)
3292 continue;
3293
3294 if (this_mode != DC_PRED) {
3295 // Disable intra modes other than DC_PRED for blocks with low variance
3296 // Threshold for intra skipping based on source variance
3297 // TODO(debargha): Specialize the threshold for super block sizes
3298 const unsigned int skip_intra_var_thresh = 64;
3299 if ((mode_search_skip_flags & FLAG_SKIP_INTRA_LOWVAR) &&
3300 x->source_variance < skip_intra_var_thresh)
3301 continue;
3302 // Only search the oblique modes if the best so far is
3303 // one of the neighboring directional modes
3304 if ((mode_search_skip_flags & FLAG_SKIP_INTRA_BESTINTER) &&
3305 (this_mode >= D45_PRED && this_mode <= TM_PRED)) {
3306 if (best_mode_index >= 0 && best_mbmode.ref_frame[0] > INTRA_FRAME)
3307 continue;
3308 }
3309 if (mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
3310 if (conditional_skipintra(this_mode, best_intra_mode)) continue;
3311 }
3312 }
3313 } else {
3314 const MV_REFERENCE_FRAME ref_frames[2] = { ref_frame, second_ref_frame };
3315 if (!check_best_zero_mv(cpi, mbmi_ext->mode_context, frame_mv, this_mode,
3316 ref_frames))
3317 continue;
3318 }
3319
3320 mi->mode = this_mode;
3321 mi->uv_mode = DC_PRED;
3322 mi->ref_frame[0] = ref_frame;
3323 mi->ref_frame[1] = second_ref_frame;
3324 // Evaluate all sub-pel filters irrespective of whether we can use
3325 // them for this frame.
3326 mi->interp_filter =
3327 cm->interp_filter == SWITCHABLE ? EIGHTTAP : cm->interp_filter;
3328 mi->mv[0].as_int = mi->mv[1].as_int = 0;
3329
3330 x->skip = 0;
3331 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
3332
3333 // Select prediction reference frames.
3334 for (i = 0; i < MAX_MB_PLANE; i++) {
3335 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
3336 if (comp_pred) xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
3337 }
3338
3339 if (ref_frame == INTRA_FRAME) {
3340 TX_SIZE uv_tx;
3341 struct macroblockd_plane *const pd = &xd->plane[1];
3342 memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
3343 super_block_yrd(cpi, x, &rate_y, &distortion_y, &skippable, NULL, bsize,
3344 best_rd);
3345 if (rate_y == INT_MAX) continue;
3346
3347 uv_tx = uv_txsize_lookup[bsize][mi->tx_size][pd->subsampling_x]
3348 [pd->subsampling_y];
3349 if (rate_uv_intra[uv_tx] == INT_MAX) {
3350 choose_intra_uv_mode(cpi, x, ctx, bsize, uv_tx, &rate_uv_intra[uv_tx],
3351 &rate_uv_tokenonly[uv_tx], &dist_uv[uv_tx],
3352 &skip_uv[uv_tx], &mode_uv[uv_tx]);
3353 }
3354
3355 rate_uv = rate_uv_tokenonly[uv_tx];
3356 distortion_uv = dist_uv[uv_tx];
3357 skippable = skippable && skip_uv[uv_tx];
3358 mi->uv_mode = mode_uv[uv_tx];
3359
3360 rate2 = rate_y + cpi->mbmode_cost[mi->mode] + rate_uv_intra[uv_tx];
3361 if (this_mode != DC_PRED && this_mode != TM_PRED)
3362 rate2 += intra_cost_penalty;
3363 distortion2 = distortion_y + distortion_uv;
3364 } else {
3365 this_rd = handle_inter_mode(
3366 cpi, x, bsize, &rate2, &distortion2, &skippable, &rate_y, &rate_uv,
3367 &disable_skip, frame_mv, mi_row, mi_col, single_newmv,
3368 single_inter_filter, single_skippable, &total_sse, best_rd,
3369 &mask_filter, filter_cache);
3370 if (this_rd == INT64_MAX) continue;
3371
3372 compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
3373
3374 if (cm->reference_mode == REFERENCE_MODE_SELECT) rate2 += compmode_cost;
3375 }
3376
3377 // Estimate the reference frame signaling cost and add it
3378 // to the rolling cost variable.
3379 if (comp_pred) {
3380 rate2 += ref_costs_comp[ref_frame];
3381 } else {
3382 rate2 += ref_costs_single[ref_frame];
3383 }
3384
3385 if (!disable_skip) {
3386 const vpx_prob skip_prob = vp9_get_skip_prob(cm, xd);
3387 const int skip_cost0 = vp9_cost_bit(skip_prob, 0);
3388 const int skip_cost1 = vp9_cost_bit(skip_prob, 1);
3389
3390 if (skippable) {
3391 // Back out the coefficient coding costs
3392 rate2 -= (rate_y + rate_uv);
3393
3394 // Cost the skip mb case
3395 rate2 += skip_cost1;
3396 } else if (ref_frame != INTRA_FRAME && !xd->lossless) {
3397 if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv + skip_cost0,
3398 distortion2) <
3399 RDCOST(x->rdmult, x->rddiv, skip_cost1, total_sse)) {
3400 // Add in the cost of the no skip flag.
3401 rate2 += skip_cost0;
3402 } else {
3403 // FIXME(rbultje) make this work for splitmv also
3404 assert(total_sse >= 0);
3405
3406 rate2 += skip_cost1;
3407 distortion2 = total_sse;
3408 rate2 -= (rate_y + rate_uv);
3409 this_skip2 = 1;
3410 }
3411 } else {
3412 // Add in the cost of the no skip flag.
3413 rate2 += skip_cost0;
3414 }
3415
3416 // Calculate the final RD estimate for this mode.
3417 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3418 }
3419
3420 // Apply an adjustment to the rd value based on the similarity of the
3421 // source variance and reconstructed variance.
3422 rd_variance_adjustment(cpi, x, bsize, &this_rd, ref_frame,
3423 x->source_variance);
3424
3425 if (ref_frame == INTRA_FRAME) {
3426 // Keep record of best intra rd
3427 if (this_rd < best_intra_rd) {
3428 best_intra_rd = this_rd;
3429 best_intra_mode = mi->mode;
3430 }
3431 }
3432
3433 if (!disable_skip && ref_frame == INTRA_FRAME) {
3434 for (i = 0; i < REFERENCE_MODES; ++i)
3435 best_pred_rd[i] = VPXMIN(best_pred_rd[i], this_rd);
3436 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3437 best_filter_rd[i] = VPXMIN(best_filter_rd[i], this_rd);
3438 }
3439
3440 // Did this mode help.. i.e. is it the new best mode
3441 if (this_rd < best_rd || x->skip) {
3442 int max_plane = MAX_MB_PLANE;
3443 if (!mode_excluded) {
3444 // Note index of best mode so far
3445 best_mode_index = mode_index;
3446
3447 if (ref_frame == INTRA_FRAME) {
3448 /* required for left and above block mv */
3449 mi->mv[0].as_int = 0;
3450 max_plane = 1;
3451 // Initialize interp_filter here so we do not have to check for
3452 // inter block modes in get_pred_context_switchable_interp()
3453 mi->interp_filter = SWITCHABLE_FILTERS;
3454 } else {
3455 best_pred_sse = x->pred_sse[ref_frame];
3456 }
3457
3458 rd_cost->rate = rate2;
3459 rd_cost->dist = distortion2;
3460 rd_cost->rdcost = this_rd;
3461 best_rd = this_rd;
3462 best_mbmode = *mi;
3463 best_skip2 = this_skip2;
3464 best_mode_skippable = skippable;
3465
3466 if (!x->select_tx_size) swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
3467 memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mi->tx_size],
3468 sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
3469
3470 // TODO(debargha): enhance this test with a better distortion prediction
3471 // based on qp, activity mask and history
3472 if ((mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
3473 (mode_index > MIN_EARLY_TERM_INDEX)) {
3474 int qstep = xd->plane[0].dequant[1];
3475 // TODO(debargha): Enhance this by specializing for each mode_index
3476 int scale = 4;
3477 #if CONFIG_VP9_HIGHBITDEPTH
3478 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
3479 qstep >>= (xd->bd - 8);
3480 }
3481 #endif // CONFIG_VP9_HIGHBITDEPTH
3482 if (x->source_variance < UINT_MAX) {
3483 const int var_adjust = (x->source_variance < 16);
3484 scale -= var_adjust;
3485 }
3486 if (ref_frame > INTRA_FRAME && distortion2 * scale < qstep * qstep) {
3487 early_term = 1;
3488 }
3489 }
3490 }
3491 }
3492
3493 /* keep record of best compound/single-only prediction */
3494 if (!disable_skip && ref_frame != INTRA_FRAME) {
3495 int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
3496
3497 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
3498 single_rate = rate2 - compmode_cost;
3499 hybrid_rate = rate2;
3500 } else {
3501 single_rate = rate2;
3502 hybrid_rate = rate2 + compmode_cost;
3503 }
3504
3505 single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
3506 hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
3507
3508 if (!comp_pred) {
3509 if (single_rd < best_pred_rd[SINGLE_REFERENCE])
3510 best_pred_rd[SINGLE_REFERENCE] = single_rd;
3511 } else {
3512 if (single_rd < best_pred_rd[COMPOUND_REFERENCE])
3513 best_pred_rd[COMPOUND_REFERENCE] = single_rd;
3514 }
3515 if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
3516 best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
3517
3518 /* keep record of best filter type */
3519 if (!mode_excluded && cm->interp_filter != BILINEAR) {
3520 int64_t ref =
3521 filter_cache[cm->interp_filter == SWITCHABLE ? SWITCHABLE_FILTERS
3522 : cm->interp_filter];
3523
3524 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3525 int64_t adj_rd;
3526 if (ref == INT64_MAX)
3527 adj_rd = 0;
3528 else if (filter_cache[i] == INT64_MAX)
3529 // when early termination is triggered, the encoder does not have
3530 // access to the rate-distortion cost. it only knows that the cost
3531 // should be above the maximum valid value. hence it takes the known
3532 // maximum plus an arbitrary constant as the rate-distortion cost.
3533 adj_rd = mask_filter - ref + 10;
3534 else
3535 adj_rd = filter_cache[i] - ref;
3536
3537 adj_rd += this_rd;
3538 best_filter_rd[i] = VPXMIN(best_filter_rd[i], adj_rd);
3539 }
3540 }
3541 }
3542
3543 if (early_term) break;
3544
3545 if (x->skip && !comp_pred) break;
3546 }
3547
3548 // The inter modes' rate costs are not calculated precisely in some cases.
3549 // Therefore, sometimes, NEWMV is chosen instead of NEARESTMV, NEARMV, and
3550 // ZEROMV. Here, checks are added for those cases, and the mode decisions
3551 // are corrected.
3552 if (best_mbmode.mode == NEWMV) {
3553 const MV_REFERENCE_FRAME refs[2] = { best_mbmode.ref_frame[0],
3554 best_mbmode.ref_frame[1] };
3555 int comp_pred_mode = refs[1] > INTRA_FRAME;
3556
3557 if (frame_mv[NEARESTMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3558 ((comp_pred_mode &&
3559 frame_mv[NEARESTMV][refs[1]].as_int == best_mbmode.mv[1].as_int) ||
3560 !comp_pred_mode))
3561 best_mbmode.mode = NEARESTMV;
3562 else if (frame_mv[NEARMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3563 ((comp_pred_mode &&
3564 frame_mv[NEARMV][refs[1]].as_int == best_mbmode.mv[1].as_int) ||
3565 !comp_pred_mode))
3566 best_mbmode.mode = NEARMV;
3567 else if (best_mbmode.mv[0].as_int == 0 &&
3568 ((comp_pred_mode && best_mbmode.mv[1].as_int == 0) ||
3569 !comp_pred_mode))
3570 best_mbmode.mode = ZEROMV;
3571 }
3572
3573 if (best_mode_index < 0 || best_rd >= best_rd_so_far) {
3574 rd_cost->rate = INT_MAX;
3575 rd_cost->rdcost = INT64_MAX;
3576 return;
3577 }
3578
3579 // If we used an estimate for the uv intra rd in the loop above...
3580 if (sf->use_uv_intra_rd_estimate) {
3581 // Do Intra UV best rd mode selection if best mode choice above was intra.
3582 if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
3583 TX_SIZE uv_tx_size;
3584 *mi = best_mbmode;
3585 uv_tx_size = get_uv_tx_size(mi, &xd->plane[1]);
3586 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra[uv_tx_size],
3587 &rate_uv_tokenonly[uv_tx_size],
3588 &dist_uv[uv_tx_size], &skip_uv[uv_tx_size],
3589 bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize,
3590 uv_tx_size);
3591 }
3592 }
3593
3594 assert((cm->interp_filter == SWITCHABLE) ||
3595 (cm->interp_filter == best_mbmode.interp_filter) ||
3596 !is_inter_block(&best_mbmode));
3597
3598 if (!cpi->rc.is_src_frame_alt_ref)
3599 vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
3600 sf->adaptive_rd_thresh, bsize, best_mode_index);
3601
3602 // macroblock modes
3603 *mi = best_mbmode;
3604 x->skip |= best_skip2;
3605
3606 for (i = 0; i < REFERENCE_MODES; ++i) {
3607 if (best_pred_rd[i] == INT64_MAX)
3608 best_pred_diff[i] = INT_MIN;
3609 else
3610 best_pred_diff[i] = best_rd - best_pred_rd[i];
3611 }
3612
3613 if (!x->skip) {
3614 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3615 if (best_filter_rd[i] == INT64_MAX)
3616 best_filter_diff[i] = 0;
3617 else
3618 best_filter_diff[i] = best_rd - best_filter_rd[i];
3619 }
3620 if (cm->interp_filter == SWITCHABLE)
3621 assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
3622 } else {
3623 vp9_zero(best_filter_diff);
3624 }
3625
3626 // TODO(yunqingwang): Moving this line in front of the above best_filter_diff
3627 // updating code causes PSNR loss. Need to figure out the confliction.
3628 x->skip |= best_mode_skippable;
3629
3630 if (!x->skip && !x->select_tx_size) {
3631 int has_high_freq_coeff = 0;
3632 int plane;
3633 int max_plane = is_inter_block(xd->mi[0]) ? MAX_MB_PLANE : 1;
3634 for (plane = 0; plane < max_plane; ++plane) {
3635 x->plane[plane].eobs = ctx->eobs_pbuf[plane][1];
3636 has_high_freq_coeff |= vp9_has_high_freq_in_plane(x, bsize, plane);
3637 }
3638
3639 for (plane = max_plane; plane < MAX_MB_PLANE; ++plane) {
3640 x->plane[plane].eobs = ctx->eobs_pbuf[plane][2];
3641 has_high_freq_coeff |= vp9_has_high_freq_in_plane(x, bsize, plane);
3642 }
3643
3644 best_mode_skippable |= !has_high_freq_coeff;
3645 }
3646
3647 assert(best_mode_index >= 0);
3648
3649 store_coding_context(x, ctx, best_mode_index, best_pred_diff,
3650 best_filter_diff, best_mode_skippable);
3651 }
3652
vp9_rd_pick_inter_mode_sb_seg_skip(VP9_COMP * cpi,TileDataEnc * tile_data,MACROBLOCK * x,RD_COST * rd_cost,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx,int64_t best_rd_so_far)3653 void vp9_rd_pick_inter_mode_sb_seg_skip(VP9_COMP *cpi, TileDataEnc *tile_data,
3654 MACROBLOCK *x, RD_COST *rd_cost,
3655 BLOCK_SIZE bsize,
3656 PICK_MODE_CONTEXT *ctx,
3657 int64_t best_rd_so_far) {
3658 VP9_COMMON *const cm = &cpi->common;
3659 MACROBLOCKD *const xd = &x->e_mbd;
3660 MODE_INFO *const mi = xd->mi[0];
3661 unsigned char segment_id = mi->segment_id;
3662 const int comp_pred = 0;
3663 int i;
3664 int64_t best_pred_diff[REFERENCE_MODES];
3665 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3666 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3667 vpx_prob comp_mode_p;
3668 INTERP_FILTER best_filter = SWITCHABLE;
3669 int64_t this_rd = INT64_MAX;
3670 int rate2 = 0;
3671 const int64_t distortion2 = 0;
3672
3673 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3674
3675 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3676 &comp_mode_p);
3677
3678 for (i = 0; i < MAX_REF_FRAMES; ++i) x->pred_sse[i] = INT_MAX;
3679 for (i = LAST_FRAME; i < MAX_REF_FRAMES; ++i) x->pred_mv_sad[i] = INT_MAX;
3680
3681 rd_cost->rate = INT_MAX;
3682
3683 assert(segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP));
3684
3685 mi->mode = ZEROMV;
3686 mi->uv_mode = DC_PRED;
3687 mi->ref_frame[0] = LAST_FRAME;
3688 mi->ref_frame[1] = NONE;
3689 mi->mv[0].as_int = 0;
3690 x->skip = 1;
3691
3692 if (cm->interp_filter != BILINEAR) {
3693 best_filter = EIGHTTAP;
3694 if (cm->interp_filter == SWITCHABLE &&
3695 x->source_variance >= cpi->sf.disable_filter_search_var_thresh) {
3696 int rs;
3697 int best_rs = INT_MAX;
3698 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
3699 mi->interp_filter = i;
3700 rs = vp9_get_switchable_rate(cpi, xd);
3701 if (rs < best_rs) {
3702 best_rs = rs;
3703 best_filter = mi->interp_filter;
3704 }
3705 }
3706 }
3707 }
3708 // Set the appropriate filter
3709 if (cm->interp_filter == SWITCHABLE) {
3710 mi->interp_filter = best_filter;
3711 rate2 += vp9_get_switchable_rate(cpi, xd);
3712 } else {
3713 mi->interp_filter = cm->interp_filter;
3714 }
3715
3716 if (cm->reference_mode == REFERENCE_MODE_SELECT)
3717 rate2 += vp9_cost_bit(comp_mode_p, comp_pred);
3718
3719 // Estimate the reference frame signaling cost and add it
3720 // to the rolling cost variable.
3721 rate2 += ref_costs_single[LAST_FRAME];
3722 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3723
3724 rd_cost->rate = rate2;
3725 rd_cost->dist = distortion2;
3726 rd_cost->rdcost = this_rd;
3727
3728 if (this_rd >= best_rd_so_far) {
3729 rd_cost->rate = INT_MAX;
3730 rd_cost->rdcost = INT64_MAX;
3731 return;
3732 }
3733
3734 assert((cm->interp_filter == SWITCHABLE) ||
3735 (cm->interp_filter == mi->interp_filter));
3736
3737 vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
3738 cpi->sf.adaptive_rd_thresh, bsize, THR_ZEROMV);
3739
3740 vp9_zero(best_pred_diff);
3741 vp9_zero(best_filter_diff);
3742
3743 if (!x->select_tx_size) swap_block_ptr(x, ctx, 1, 0, 0, MAX_MB_PLANE);
3744 store_coding_context(x, ctx, THR_ZEROMV, best_pred_diff, best_filter_diff, 0);
3745 }
3746
vp9_rd_pick_inter_mode_sub8x8(VP9_COMP * cpi,TileDataEnc * tile_data,MACROBLOCK * x,int mi_row,int mi_col,RD_COST * rd_cost,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx,int64_t best_rd_so_far)3747 void vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, TileDataEnc *tile_data,
3748 MACROBLOCK *x, int mi_row, int mi_col,
3749 RD_COST *rd_cost, BLOCK_SIZE bsize,
3750 PICK_MODE_CONTEXT *ctx,
3751 int64_t best_rd_so_far) {
3752 VP9_COMMON *const cm = &cpi->common;
3753 RD_OPT *const rd_opt = &cpi->rd;
3754 SPEED_FEATURES *const sf = &cpi->sf;
3755 MACROBLOCKD *const xd = &x->e_mbd;
3756 MODE_INFO *const mi = xd->mi[0];
3757 const struct segmentation *const seg = &cm->seg;
3758 MV_REFERENCE_FRAME ref_frame, second_ref_frame;
3759 unsigned char segment_id = mi->segment_id;
3760 int comp_pred, i;
3761 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
3762 struct buf_2d yv12_mb[4][MAX_MB_PLANE];
3763 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
3764 VP9_ALT_FLAG };
3765 int64_t best_rd = best_rd_so_far;
3766 int64_t best_yrd = best_rd_so_far; // FIXME(rbultje) more precise
3767 int64_t best_pred_diff[REFERENCE_MODES];
3768 int64_t best_pred_rd[REFERENCE_MODES];
3769 int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
3770 int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3771 MODE_INFO best_mbmode;
3772 int ref_index, best_ref_index = 0;
3773 unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3774 vpx_prob comp_mode_p;
3775 INTERP_FILTER tmp_best_filter = SWITCHABLE;
3776 int rate_uv_intra, rate_uv_tokenonly;
3777 int64_t dist_uv;
3778 int skip_uv;
3779 PREDICTION_MODE mode_uv = DC_PRED;
3780 const int intra_cost_penalty = vp9_get_intra_cost_penalty(
3781 cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
3782 int_mv seg_mvs[4][MAX_REF_FRAMES];
3783 b_mode_info best_bmodes[4];
3784 int best_skip2 = 0;
3785 int ref_frame_skip_mask[2] = { 0 };
3786 int64_t mask_filter = 0;
3787 int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
3788 int internal_active_edge =
3789 vp9_active_edge_sb(cpi, mi_row, mi_col) && vp9_internal_image_edge(cpi);
3790
3791 x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3792 memset(x->zcoeff_blk[TX_4X4], 0, 4);
3793 vp9_zero(best_mbmode);
3794
3795 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i) filter_cache[i] = INT64_MAX;
3796
3797 for (i = 0; i < 4; i++) {
3798 int j;
3799 for (j = 0; j < MAX_REF_FRAMES; j++) seg_mvs[i][j].as_int = INVALID_MV;
3800 }
3801
3802 estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3803 &comp_mode_p);
3804
3805 for (i = 0; i < REFERENCE_MODES; ++i) best_pred_rd[i] = INT64_MAX;
3806 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3807 best_filter_rd[i] = INT64_MAX;
3808 rate_uv_intra = INT_MAX;
3809
3810 rd_cost->rate = INT_MAX;
3811
3812 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ref_frame++) {
3813 if (cpi->ref_frame_flags & flag_list[ref_frame]) {
3814 setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
3815 frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
3816 } else {
3817 ref_frame_skip_mask[0] |= (1 << ref_frame);
3818 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3819 }
3820 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
3821 frame_mv[ZEROMV][ref_frame].as_int = 0;
3822 }
3823
3824 for (ref_index = 0; ref_index < MAX_REFS; ++ref_index) {
3825 int mode_excluded = 0;
3826 int64_t this_rd = INT64_MAX;
3827 int disable_skip = 0;
3828 int compmode_cost = 0;
3829 int rate2 = 0, rate_y = 0, rate_uv = 0;
3830 int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
3831 int skippable = 0;
3832 int i;
3833 int this_skip2 = 0;
3834 int64_t total_sse = INT_MAX;
3835 int early_term = 0;
3836 struct buf_2d backup_yv12[2][MAX_MB_PLANE];
3837
3838 ref_frame = vp9_ref_order[ref_index].ref_frame[0];
3839 second_ref_frame = vp9_ref_order[ref_index].ref_frame[1];
3840
3841 #if CONFIG_BETTER_HW_COMPATIBILITY
3842 // forbid 8X4 and 4X8 partitions if any reference frame is scaled.
3843 if (bsize == BLOCK_8X4 || bsize == BLOCK_4X8) {
3844 int ref_scaled = vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf);
3845 if (second_ref_frame > INTRA_FRAME)
3846 ref_scaled += vp9_is_scaled(&cm->frame_refs[second_ref_frame - 1].sf);
3847 if (ref_scaled) continue;
3848 }
3849 #endif
3850 // Look at the reference frame of the best mode so far and set the
3851 // skip mask to look at a subset of the remaining modes.
3852 if (ref_index > 2 && sf->mode_skip_start < MAX_MODES) {
3853 if (ref_index == 3) {
3854 switch (best_mbmode.ref_frame[0]) {
3855 case INTRA_FRAME: break;
3856 case LAST_FRAME:
3857 ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME);
3858 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3859 break;
3860 case GOLDEN_FRAME:
3861 ref_frame_skip_mask[0] |= (1 << LAST_FRAME) | (1 << ALTREF_FRAME);
3862 ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3863 break;
3864 case ALTREF_FRAME:
3865 ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << LAST_FRAME);
3866 break;
3867 case NONE:
3868 case MAX_REF_FRAMES: assert(0 && "Invalid Reference frame"); break;
3869 }
3870 }
3871 }
3872
3873 if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
3874 (ref_frame_skip_mask[1] & (1 << VPXMAX(0, second_ref_frame))))
3875 continue;
3876
3877 // Test best rd so far against threshold for trying this mode.
3878 if (!internal_active_edge &&
3879 rd_less_than_thresh(best_rd,
3880 rd_opt->threshes[segment_id][bsize][ref_index],
3881 tile_data->thresh_freq_fact[bsize][ref_index]))
3882 continue;
3883
3884 comp_pred = second_ref_frame > INTRA_FRAME;
3885 if (comp_pred) {
3886 if (!cpi->allow_comp_inter_inter) continue;
3887 if (!(cpi->ref_frame_flags & flag_list[second_ref_frame])) continue;
3888 // Do not allow compound prediction if the segment level reference frame
3889 // feature is in use as in this case there can only be one reference.
3890 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) continue;
3891
3892 if ((sf->mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
3893 best_mbmode.ref_frame[0] == INTRA_FRAME)
3894 continue;
3895 }
3896
3897 if (comp_pred)
3898 mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
3899 else if (ref_frame != INTRA_FRAME)
3900 mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
3901
3902 // If the segment reference frame feature is enabled....
3903 // then do nothing if the current ref frame is not allowed..
3904 if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
3905 get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
3906 continue;
3907 // Disable this drop out case if the ref frame
3908 // segment level feature is enabled for this segment. This is to
3909 // prevent the possibility that we end up unable to pick any mode.
3910 } else if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
3911 // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
3912 // unless ARNR filtering is enabled in which case we want
3913 // an unfiltered alternative. We allow near/nearest as well
3914 // because they may result in zero-zero MVs but be cheaper.
3915 if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0))
3916 continue;
3917 }
3918
3919 mi->tx_size = TX_4X4;
3920 mi->uv_mode = DC_PRED;
3921 mi->ref_frame[0] = ref_frame;
3922 mi->ref_frame[1] = second_ref_frame;
3923 // Evaluate all sub-pel filters irrespective of whether we can use
3924 // them for this frame.
3925 mi->interp_filter =
3926 cm->interp_filter == SWITCHABLE ? EIGHTTAP : cm->interp_filter;
3927 x->skip = 0;
3928 set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
3929
3930 // Select prediction reference frames.
3931 for (i = 0; i < MAX_MB_PLANE; i++) {
3932 xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
3933 if (comp_pred) xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
3934 }
3935
3936 if (ref_frame == INTRA_FRAME) {
3937 int rate;
3938 if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate, &rate_y, &distortion_y,
3939 best_rd) >= best_rd)
3940 continue;
3941 rate2 += rate;
3942 rate2 += intra_cost_penalty;
3943 distortion2 += distortion_y;
3944
3945 if (rate_uv_intra == INT_MAX) {
3946 choose_intra_uv_mode(cpi, x, ctx, bsize, TX_4X4, &rate_uv_intra,
3947 &rate_uv_tokenonly, &dist_uv, &skip_uv, &mode_uv);
3948 }
3949 rate2 += rate_uv_intra;
3950 rate_uv = rate_uv_tokenonly;
3951 distortion2 += dist_uv;
3952 distortion_uv = dist_uv;
3953 mi->uv_mode = mode_uv;
3954 } else {
3955 int rate;
3956 int64_t distortion;
3957 int64_t this_rd_thresh;
3958 int64_t tmp_rd, tmp_best_rd = INT64_MAX, tmp_best_rdu = INT64_MAX;
3959 int tmp_best_rate = INT_MAX, tmp_best_ratey = INT_MAX;
3960 int64_t tmp_best_distortion = INT_MAX, tmp_best_sse, uv_sse;
3961 int tmp_best_skippable = 0;
3962 int switchable_filter_index;
3963 int_mv *second_ref =
3964 comp_pred ? &x->mbmi_ext->ref_mvs[second_ref_frame][0] : NULL;
3965 b_mode_info tmp_best_bmodes[16];
3966 MODE_INFO tmp_best_mbmode;
3967 BEST_SEG_INFO bsi[SWITCHABLE_FILTERS];
3968 int pred_exists = 0;
3969 int uv_skippable;
3970
3971 YV12_BUFFER_CONFIG *scaled_ref_frame[2] = { NULL, NULL };
3972 int ref;
3973
3974 for (ref = 0; ref < 2; ++ref) {
3975 scaled_ref_frame[ref] =
3976 mi->ref_frame[ref] > INTRA_FRAME
3977 ? vp9_get_scaled_ref_frame(cpi, mi->ref_frame[ref])
3978 : NULL;
3979
3980 if (scaled_ref_frame[ref]) {
3981 int i;
3982 // Swap out the reference frame for a version that's been scaled to
3983 // match the resolution of the current frame, allowing the existing
3984 // motion search code to be used without additional modifications.
3985 for (i = 0; i < MAX_MB_PLANE; i++)
3986 backup_yv12[ref][i] = xd->plane[i].pre[ref];
3987 vp9_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
3988 NULL);
3989 }
3990 }
3991
3992 this_rd_thresh = (ref_frame == LAST_FRAME)
3993 ? rd_opt->threshes[segment_id][bsize][THR_LAST]
3994 : rd_opt->threshes[segment_id][bsize][THR_ALTR];
3995 this_rd_thresh = (ref_frame == GOLDEN_FRAME)
3996 ? rd_opt->threshes[segment_id][bsize][THR_GOLD]
3997 : this_rd_thresh;
3998 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
3999 filter_cache[i] = INT64_MAX;
4000
4001 if (cm->interp_filter != BILINEAR) {
4002 tmp_best_filter = EIGHTTAP;
4003 if (x->source_variance < sf->disable_filter_search_var_thresh) {
4004 tmp_best_filter = EIGHTTAP;
4005 } else if (sf->adaptive_pred_interp_filter == 1 &&
4006 ctx->pred_interp_filter < SWITCHABLE) {
4007 tmp_best_filter = ctx->pred_interp_filter;
4008 } else if (sf->adaptive_pred_interp_filter == 2) {
4009 tmp_best_filter = ctx->pred_interp_filter < SWITCHABLE
4010 ? ctx->pred_interp_filter
4011 : 0;
4012 } else {
4013 for (switchable_filter_index = 0;
4014 switchable_filter_index < SWITCHABLE_FILTERS;
4015 ++switchable_filter_index) {
4016 int newbest, rs;
4017 int64_t rs_rd;
4018 MB_MODE_INFO_EXT *mbmi_ext = x->mbmi_ext;
4019 mi->interp_filter = switchable_filter_index;
4020 tmp_rd = rd_pick_best_sub8x8_mode(
4021 cpi, x, &mbmi_ext->ref_mvs[ref_frame][0], second_ref, best_yrd,
4022 &rate, &rate_y, &distortion, &skippable, &total_sse,
4023 (int)this_rd_thresh, seg_mvs, bsi, switchable_filter_index,
4024 mi_row, mi_col);
4025
4026 if (tmp_rd == INT64_MAX) continue;
4027 rs = vp9_get_switchable_rate(cpi, xd);
4028 rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
4029 filter_cache[switchable_filter_index] = tmp_rd;
4030 filter_cache[SWITCHABLE_FILTERS] =
4031 VPXMIN(filter_cache[SWITCHABLE_FILTERS], tmp_rd + rs_rd);
4032 if (cm->interp_filter == SWITCHABLE) tmp_rd += rs_rd;
4033
4034 mask_filter = VPXMAX(mask_filter, tmp_rd);
4035
4036 newbest = (tmp_rd < tmp_best_rd);
4037 if (newbest) {
4038 tmp_best_filter = mi->interp_filter;
4039 tmp_best_rd = tmp_rd;
4040 }
4041 if ((newbest && cm->interp_filter == SWITCHABLE) ||
4042 (mi->interp_filter == cm->interp_filter &&
4043 cm->interp_filter != SWITCHABLE)) {
4044 tmp_best_rdu = tmp_rd;
4045 tmp_best_rate = rate;
4046 tmp_best_ratey = rate_y;
4047 tmp_best_distortion = distortion;
4048 tmp_best_sse = total_sse;
4049 tmp_best_skippable = skippable;
4050 tmp_best_mbmode = *mi;
4051 for (i = 0; i < 4; i++) {
4052 tmp_best_bmodes[i] = xd->mi[0]->bmi[i];
4053 x->zcoeff_blk[TX_4X4][i] = !x->plane[0].eobs[i];
4054 }
4055 pred_exists = 1;
4056 if (switchable_filter_index == 0 && sf->use_rd_breakout &&
4057 best_rd < INT64_MAX) {
4058 if (tmp_best_rdu / 2 > best_rd) {
4059 // skip searching the other filters if the first is
4060 // already substantially larger than the best so far
4061 tmp_best_filter = mi->interp_filter;
4062 tmp_best_rdu = INT64_MAX;
4063 break;
4064 }
4065 }
4066 }
4067 } // switchable_filter_index loop
4068 }
4069 }
4070
4071 if (tmp_best_rdu == INT64_MAX && pred_exists) continue;
4072
4073 mi->interp_filter = (cm->interp_filter == SWITCHABLE ? tmp_best_filter
4074 : cm->interp_filter);
4075 if (!pred_exists) {
4076 // Handles the special case when a filter that is not in the
4077 // switchable list (bilinear, 6-tap) is indicated at the frame level
4078 tmp_rd = rd_pick_best_sub8x8_mode(
4079 cpi, x, &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, 0, mi_row, mi_col);
4082 if (tmp_rd == INT64_MAX) continue;
4083 } else {
4084 total_sse = tmp_best_sse;
4085 rate = tmp_best_rate;
4086 rate_y = tmp_best_ratey;
4087 distortion = tmp_best_distortion;
4088 skippable = tmp_best_skippable;
4089 *mi = tmp_best_mbmode;
4090 for (i = 0; i < 4; i++) xd->mi[0]->bmi[i] = tmp_best_bmodes[i];
4091 }
4092
4093 rate2 += rate;
4094 distortion2 += distortion;
4095
4096 if (cm->interp_filter == SWITCHABLE)
4097 rate2 += vp9_get_switchable_rate(cpi, xd);
4098
4099 if (!mode_excluded)
4100 mode_excluded = comp_pred ? cm->reference_mode == SINGLE_REFERENCE
4101 : cm->reference_mode == COMPOUND_REFERENCE;
4102
4103 compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
4104
4105 tmp_best_rdu =
4106 best_rd - VPXMIN(RDCOST(x->rdmult, x->rddiv, rate2, distortion2),
4107 RDCOST(x->rdmult, x->rddiv, 0, total_sse));
4108
4109 if (tmp_best_rdu > 0) {
4110 // If even the 'Y' rd value of split is higher than best so far
4111 // then dont bother looking at UV
4112 vp9_build_inter_predictors_sbuv(&x->e_mbd, mi_row, mi_col, BLOCK_8X8);
4113 memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
4114 if (!super_block_uvrd(cpi, x, &rate_uv, &distortion_uv, &uv_skippable,
4115 &uv_sse, BLOCK_8X8, tmp_best_rdu)) {
4116 for (ref = 0; ref < 2; ++ref) {
4117 if (scaled_ref_frame[ref]) {
4118 int i;
4119 for (i = 0; i < MAX_MB_PLANE; ++i)
4120 xd->plane[i].pre[ref] = backup_yv12[ref][i];
4121 }
4122 }
4123 continue;
4124 }
4125
4126 rate2 += rate_uv;
4127 distortion2 += distortion_uv;
4128 skippable = skippable && uv_skippable;
4129 total_sse += uv_sse;
4130 }
4131
4132 for (ref = 0; ref < 2; ++ref) {
4133 if (scaled_ref_frame[ref]) {
4134 // Restore the prediction frame pointers to their unscaled versions.
4135 int i;
4136 for (i = 0; i < MAX_MB_PLANE; ++i)
4137 xd->plane[i].pre[ref] = backup_yv12[ref][i];
4138 }
4139 }
4140 }
4141
4142 if (cm->reference_mode == REFERENCE_MODE_SELECT) rate2 += compmode_cost;
4143
4144 // Estimate the reference frame signaling cost and add it
4145 // to the rolling cost variable.
4146 if (second_ref_frame > INTRA_FRAME) {
4147 rate2 += ref_costs_comp[ref_frame];
4148 } else {
4149 rate2 += ref_costs_single[ref_frame];
4150 }
4151
4152 if (!disable_skip) {
4153 const vpx_prob skip_prob = vp9_get_skip_prob(cm, xd);
4154 const int skip_cost0 = vp9_cost_bit(skip_prob, 0);
4155 const int skip_cost1 = vp9_cost_bit(skip_prob, 1);
4156
4157 // Skip is never coded at the segment level for sub8x8 blocks and instead
4158 // always coded in the bitstream at the mode info level.
4159 if (ref_frame != INTRA_FRAME && !xd->lossless) {
4160 if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv + skip_cost0,
4161 distortion2) <
4162 RDCOST(x->rdmult, x->rddiv, skip_cost1, total_sse)) {
4163 // Add in the cost of the no skip flag.
4164 rate2 += skip_cost0;
4165 } else {
4166 // FIXME(rbultje) make this work for splitmv also
4167 rate2 += skip_cost1;
4168 distortion2 = total_sse;
4169 assert(total_sse >= 0);
4170 rate2 -= (rate_y + rate_uv);
4171 rate_y = 0;
4172 rate_uv = 0;
4173 this_skip2 = 1;
4174 }
4175 } else {
4176 // Add in the cost of the no skip flag.
4177 rate2 += skip_cost0;
4178 }
4179
4180 // Calculate the final RD estimate for this mode.
4181 this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
4182 }
4183
4184 if (!disable_skip && ref_frame == INTRA_FRAME) {
4185 for (i = 0; i < REFERENCE_MODES; ++i)
4186 best_pred_rd[i] = VPXMIN(best_pred_rd[i], this_rd);
4187 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
4188 best_filter_rd[i] = VPXMIN(best_filter_rd[i], this_rd);
4189 }
4190
4191 // Did this mode help.. i.e. is it the new best mode
4192 if (this_rd < best_rd || x->skip) {
4193 if (!mode_excluded) {
4194 int max_plane = MAX_MB_PLANE;
4195 // Note index of best mode so far
4196 best_ref_index = ref_index;
4197
4198 if (ref_frame == INTRA_FRAME) {
4199 /* required for left and above block mv */
4200 mi->mv[0].as_int = 0;
4201 max_plane = 1;
4202 // Initialize interp_filter here so we do not have to check for
4203 // inter block modes in get_pred_context_switchable_interp()
4204 mi->interp_filter = SWITCHABLE_FILTERS;
4205 }
4206
4207 rd_cost->rate = rate2;
4208 rd_cost->dist = distortion2;
4209 rd_cost->rdcost = this_rd;
4210 best_rd = this_rd;
4211 best_yrd =
4212 best_rd - RDCOST(x->rdmult, x->rddiv, rate_uv, distortion_uv);
4213 best_mbmode = *mi;
4214 best_skip2 = this_skip2;
4215 if (!x->select_tx_size) swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
4216 memcpy(ctx->zcoeff_blk, x->zcoeff_blk[TX_4X4],
4217 sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
4218
4219 for (i = 0; i < 4; i++) best_bmodes[i] = xd->mi[0]->bmi[i];
4220
4221 // TODO(debargha): enhance this test with a better distortion prediction
4222 // based on qp, activity mask and history
4223 if ((sf->mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
4224 (ref_index > MIN_EARLY_TERM_INDEX)) {
4225 int qstep = xd->plane[0].dequant[1];
4226 // TODO(debargha): Enhance this by specializing for each mode_index
4227 int scale = 4;
4228 #if CONFIG_VP9_HIGHBITDEPTH
4229 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
4230 qstep >>= (xd->bd - 8);
4231 }
4232 #endif // CONFIG_VP9_HIGHBITDEPTH
4233 if (x->source_variance < UINT_MAX) {
4234 const int var_adjust = (x->source_variance < 16);
4235 scale -= var_adjust;
4236 }
4237 if (ref_frame > INTRA_FRAME && distortion2 * scale < qstep * qstep) {
4238 early_term = 1;
4239 }
4240 }
4241 }
4242 }
4243
4244 /* keep record of best compound/single-only prediction */
4245 if (!disable_skip && ref_frame != INTRA_FRAME) {
4246 int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
4247
4248 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
4249 single_rate = rate2 - compmode_cost;
4250 hybrid_rate = rate2;
4251 } else {
4252 single_rate = rate2;
4253 hybrid_rate = rate2 + compmode_cost;
4254 }
4255
4256 single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
4257 hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
4258
4259 if (!comp_pred && single_rd < best_pred_rd[SINGLE_REFERENCE])
4260 best_pred_rd[SINGLE_REFERENCE] = single_rd;
4261 else if (comp_pred && single_rd < best_pred_rd[COMPOUND_REFERENCE])
4262 best_pred_rd[COMPOUND_REFERENCE] = single_rd;
4263
4264 if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
4265 best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
4266 }
4267
4268 /* keep record of best filter type */
4269 if (!mode_excluded && !disable_skip && ref_frame != INTRA_FRAME &&
4270 cm->interp_filter != BILINEAR) {
4271 int64_t ref =
4272 filter_cache[cm->interp_filter == SWITCHABLE ? SWITCHABLE_FILTERS
4273 : cm->interp_filter];
4274 int64_t adj_rd;
4275 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4276 if (ref == INT64_MAX)
4277 adj_rd = 0;
4278 else if (filter_cache[i] == INT64_MAX)
4279 // when early termination is triggered, the encoder does not have
4280 // access to the rate-distortion cost. it only knows that the cost
4281 // should be above the maximum valid value. hence it takes the known
4282 // maximum plus an arbitrary constant as the rate-distortion cost.
4283 adj_rd = mask_filter - ref + 10;
4284 else
4285 adj_rd = filter_cache[i] - ref;
4286
4287 adj_rd += this_rd;
4288 best_filter_rd[i] = VPXMIN(best_filter_rd[i], adj_rd);
4289 }
4290 }
4291
4292 if (early_term) break;
4293
4294 if (x->skip && !comp_pred) break;
4295 }
4296
4297 if (best_rd >= best_rd_so_far) {
4298 rd_cost->rate = INT_MAX;
4299 rd_cost->rdcost = INT64_MAX;
4300 return;
4301 }
4302
4303 // If we used an estimate for the uv intra rd in the loop above...
4304 if (sf->use_uv_intra_rd_estimate) {
4305 // Do Intra UV best rd mode selection if best mode choice above was intra.
4306 if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
4307 *mi = best_mbmode;
4308 rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra, &rate_uv_tokenonly,
4309 &dist_uv, &skip_uv, BLOCK_8X8, TX_4X4);
4310 }
4311 }
4312
4313 if (best_rd == INT64_MAX) {
4314 rd_cost->rate = INT_MAX;
4315 rd_cost->dist = INT64_MAX;
4316 rd_cost->rdcost = INT64_MAX;
4317 return;
4318 }
4319
4320 assert((cm->interp_filter == SWITCHABLE) ||
4321 (cm->interp_filter == best_mbmode.interp_filter) ||
4322 !is_inter_block(&best_mbmode));
4323
4324 vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact, sf->adaptive_rd_thresh,
4325 bsize, best_ref_index);
4326
4327 // macroblock modes
4328 *mi = best_mbmode;
4329 x->skip |= best_skip2;
4330 if (!is_inter_block(&best_mbmode)) {
4331 for (i = 0; i < 4; i++) xd->mi[0]->bmi[i].as_mode = best_bmodes[i].as_mode;
4332 } else {
4333 for (i = 0; i < 4; ++i)
4334 memcpy(&xd->mi[0]->bmi[i], &best_bmodes[i], sizeof(b_mode_info));
4335
4336 mi->mv[0].as_int = xd->mi[0]->bmi[3].as_mv[0].as_int;
4337 mi->mv[1].as_int = xd->mi[0]->bmi[3].as_mv[1].as_int;
4338 }
4339
4340 for (i = 0; i < REFERENCE_MODES; ++i) {
4341 if (best_pred_rd[i] == INT64_MAX)
4342 best_pred_diff[i] = INT_MIN;
4343 else
4344 best_pred_diff[i] = best_rd - best_pred_rd[i];
4345 }
4346
4347 if (!x->skip) {
4348 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4349 if (best_filter_rd[i] == INT64_MAX)
4350 best_filter_diff[i] = 0;
4351 else
4352 best_filter_diff[i] = best_rd - best_filter_rd[i];
4353 }
4354 if (cm->interp_filter == SWITCHABLE)
4355 assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
4356 } else {
4357 vp9_zero(best_filter_diff);
4358 }
4359
4360 store_coding_context(x, ctx, best_ref_index, best_pred_diff, best_filter_diff,
4361 0);
4362 }
4363