Line data Source code
1 : /*
2 : * Copyright (c) 2014 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 <limits.h>
13 : #include <math.h>
14 : #include <stdio.h>
15 :
16 : #include "./vp9_rtcd.h"
17 : #include "./vpx_dsp_rtcd.h"
18 :
19 : #include "vpx/vpx_codec.h"
20 : #include "vpx_dsp/vpx_dsp_common.h"
21 : #include "vpx_mem/vpx_mem.h"
22 : #include "vpx_ports/mem.h"
23 :
24 : #include "vp9/common/vp9_blockd.h"
25 : #include "vp9/common/vp9_common.h"
26 : #include "vp9/common/vp9_mvref_common.h"
27 : #include "vp9/common/vp9_pred_common.h"
28 : #include "vp9/common/vp9_reconinter.h"
29 : #include "vp9/common/vp9_reconintra.h"
30 : #include "vp9/common/vp9_scan.h"
31 :
32 : #include "vp9/encoder/vp9_cost.h"
33 : #include "vp9/encoder/vp9_encoder.h"
34 : #include "vp9/encoder/vp9_pickmode.h"
35 : #include "vp9/encoder/vp9_ratectrl.h"
36 : #include "vp9/encoder/vp9_rd.h"
37 :
38 : typedef struct {
39 : uint8_t *data;
40 : int stride;
41 : int in_use;
42 : } PRED_BUFFER;
43 :
44 : static const int pos_shift_16x16[4][4] = {
45 : { 9, 10, 13, 14 }, { 11, 12, 15, 16 }, { 17, 18, 21, 22 }, { 19, 20, 23, 24 }
46 : };
47 :
48 0 : static int mv_refs_rt(VP9_COMP *cpi, const VP9_COMMON *cm, const MACROBLOCK *x,
49 : const MACROBLOCKD *xd, const TileInfo *const tile,
50 : MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
51 : int_mv *mv_ref_list, int_mv *base_mv, int mi_row,
52 : int mi_col, int use_base_mv) {
53 0 : const int *ref_sign_bias = cm->ref_frame_sign_bias;
54 0 : int i, refmv_count = 0;
55 :
56 0 : const POSITION *const mv_ref_search = mv_ref_blocks[mi->sb_type];
57 :
58 0 : int different_ref_found = 0;
59 0 : int context_counter = 0;
60 0 : int const_motion = 0;
61 :
62 : // Blank the reference vector list
63 0 : memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES);
64 :
65 : // The nearest 2 blocks are treated differently
66 : // if the size < 8x8 we get the mv from the bmi substructure,
67 : // and we also need to keep a mode count.
68 0 : for (i = 0; i < 2; ++i) {
69 0 : const POSITION *const mv_ref = &mv_ref_search[i];
70 0 : if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
71 0 : const MODE_INFO *const candidate_mi =
72 0 : xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
73 : // Keep counts for entropy encoding.
74 0 : context_counter += mode_2_counter[candidate_mi->mode];
75 0 : different_ref_found = 1;
76 :
77 0 : if (candidate_mi->ref_frame[0] == ref_frame)
78 0 : ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, -1),
79 : refmv_count, mv_ref_list, Done);
80 : }
81 : }
82 :
83 0 : const_motion = 1;
84 :
85 : // Check the rest of the neighbors in much the same way
86 : // as before except we don't need to keep track of sub blocks or
87 : // mode counts.
88 0 : for (; i < MVREF_NEIGHBOURS && !refmv_count; ++i) {
89 0 : const POSITION *const mv_ref = &mv_ref_search[i];
90 0 : if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
91 0 : const MODE_INFO *const candidate_mi =
92 0 : xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
93 0 : different_ref_found = 1;
94 :
95 0 : if (candidate_mi->ref_frame[0] == ref_frame)
96 0 : ADD_MV_REF_LIST(candidate_mi->mv[0], refmv_count, mv_ref_list, Done);
97 : }
98 : }
99 :
100 : // Since we couldn't find 2 mvs from the same reference frame
101 : // go back through the neighbors and find motion vectors from
102 : // different reference frames.
103 0 : if (different_ref_found && !refmv_count) {
104 0 : for (i = 0; i < MVREF_NEIGHBOURS; ++i) {
105 0 : const POSITION *mv_ref = &mv_ref_search[i];
106 0 : if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
107 0 : const MODE_INFO *const candidate_mi =
108 0 : xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
109 :
110 : // If the candidate is INTRA we don't want to consider its mv.
111 0 : IF_DIFF_REF_FRAME_ADD_MV(candidate_mi, ref_frame, ref_sign_bias,
112 : refmv_count, mv_ref_list, Done);
113 : }
114 : }
115 : }
116 0 : if (use_base_mv &&
117 0 : !cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame &&
118 : ref_frame == LAST_FRAME) {
119 : // Get base layer mv.
120 0 : MV_REF *candidate =
121 0 : &cm->prev_frame
122 0 : ->mvs[(mi_col >> 1) + (mi_row >> 1) * (cm->mi_cols >> 1)];
123 0 : if (candidate->mv[0].as_int != INVALID_MV) {
124 0 : base_mv->as_mv.row = (candidate->mv[0].as_mv.row * 2);
125 0 : base_mv->as_mv.col = (candidate->mv[0].as_mv.col * 2);
126 0 : clamp_mv_ref(&base_mv->as_mv, xd);
127 : } else {
128 0 : base_mv->as_int = INVALID_MV;
129 : }
130 : }
131 :
132 : Done:
133 :
134 0 : x->mbmi_ext->mode_context[ref_frame] = counter_to_context[context_counter];
135 :
136 : // Clamp vectors
137 0 : for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i)
138 0 : clamp_mv_ref(&mv_ref_list[i].as_mv, xd);
139 :
140 0 : return const_motion;
141 : }
142 :
143 0 : static int combined_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
144 : BLOCK_SIZE bsize, int mi_row, int mi_col,
145 : int_mv *tmp_mv, int *rate_mv,
146 : int64_t best_rd_sofar, int use_base_mv) {
147 0 : MACROBLOCKD *xd = &x->e_mbd;
148 0 : MODE_INFO *mi = xd->mi[0];
149 0 : struct buf_2d backup_yv12[MAX_MB_PLANE] = { { 0, 0 } };
150 0 : const int step_param = cpi->sf.mv.fullpel_search_step_param;
151 0 : const int sadpb = x->sadperbit16;
152 : MV mvp_full;
153 0 : const int ref = mi->ref_frame[0];
154 0 : const MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
155 : MV center_mv;
156 : uint32_t dis;
157 : int rate_mode;
158 0 : const MvLimits tmp_mv_limits = x->mv_limits;
159 0 : int rv = 0;
160 : int cost_list[5];
161 0 : const YV12_BUFFER_CONFIG *scaled_ref_frame =
162 : vp9_get_scaled_ref_frame(cpi, ref);
163 0 : if (scaled_ref_frame) {
164 : int i;
165 : // Swap out the reference frame for a version that's been scaled to
166 : // match the resolution of the current frame, allowing the existing
167 : // motion search code to be used without additional modifications.
168 0 : for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0];
169 0 : vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
170 : }
171 0 : vp9_set_mv_search_range(&x->mv_limits, &ref_mv);
172 :
173 0 : assert(x->mv_best_ref_index[ref] <= 2);
174 0 : if (x->mv_best_ref_index[ref] < 2)
175 0 : mvp_full = x->mbmi_ext->ref_mvs[ref][x->mv_best_ref_index[ref]].as_mv;
176 : else
177 0 : mvp_full = x->pred_mv[ref];
178 :
179 0 : mvp_full.col >>= 3;
180 0 : mvp_full.row >>= 3;
181 :
182 0 : if (!use_base_mv)
183 0 : center_mv = ref_mv;
184 : else
185 0 : center_mv = tmp_mv->as_mv;
186 :
187 0 : vp9_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb,
188 : cond_cost_list(cpi, cost_list), ¢er_mv,
189 : &tmp_mv->as_mv, INT_MAX, 0);
190 :
191 0 : x->mv_limits = tmp_mv_limits;
192 :
193 : // calculate the bit cost on motion vector
194 0 : mvp_full.row = tmp_mv->as_mv.row * 8;
195 0 : mvp_full.col = tmp_mv->as_mv.col * 8;
196 :
197 0 : *rate_mv = vp9_mv_bit_cost(&mvp_full, &ref_mv, x->nmvjointcost, x->mvcost,
198 : MV_COST_WEIGHT);
199 :
200 0 : rate_mode =
201 0 : cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref]][INTER_OFFSET(NEWMV)];
202 0 : rv =
203 0 : !(RDCOST(x->rdmult, x->rddiv, (*rate_mv + rate_mode), 0) > best_rd_sofar);
204 :
205 0 : if (rv) {
206 0 : const int subpel_force_stop = use_base_mv && cpi->sf.base_mv_aggressive
207 : ? 2
208 0 : : cpi->sf.mv.subpel_force_stop;
209 0 : cpi->find_fractional_mv_step(
210 : x, &tmp_mv->as_mv, &ref_mv, cpi->common.allow_high_precision_mv,
211 0 : x->errorperbit, &cpi->fn_ptr[bsize], subpel_force_stop,
212 : cpi->sf.mv.subpel_iters_per_step, cond_cost_list(cpi, cost_list),
213 0 : x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref], NULL, 0, 0);
214 0 : *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv, x->nmvjointcost,
215 : x->mvcost, MV_COST_WEIGHT);
216 : }
217 :
218 0 : if (scaled_ref_frame) {
219 : int i;
220 0 : for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
221 : }
222 0 : return rv;
223 : }
224 :
225 0 : static void block_variance(const uint8_t *src, int src_stride,
226 : const uint8_t *ref, int ref_stride, int w, int h,
227 : unsigned int *sse, int *sum, int block_size,
228 : #if CONFIG_VP9_HIGHBITDEPTH
229 : int use_highbitdepth, vpx_bit_depth_t bd,
230 : #endif
231 : uint32_t *sse8x8, int *sum8x8, uint32_t *var8x8) {
232 0 : int i, j, k = 0;
233 :
234 0 : *sse = 0;
235 0 : *sum = 0;
236 :
237 0 : for (i = 0; i < h; i += block_size) {
238 0 : for (j = 0; j < w; j += block_size) {
239 : #if CONFIG_VP9_HIGHBITDEPTH
240 : if (use_highbitdepth) {
241 : switch (bd) {
242 : case VPX_BITS_8:
243 : vpx_highbd_8_get8x8var(src + src_stride * i + j, src_stride,
244 : ref + ref_stride * i + j, ref_stride,
245 : &sse8x8[k], &sum8x8[k]);
246 : break;
247 : case VPX_BITS_10:
248 : vpx_highbd_10_get8x8var(src + src_stride * i + j, src_stride,
249 : ref + ref_stride * i + j, ref_stride,
250 : &sse8x8[k], &sum8x8[k]);
251 : break;
252 : case VPX_BITS_12:
253 : vpx_highbd_12_get8x8var(src + src_stride * i + j, src_stride,
254 : ref + ref_stride * i + j, ref_stride,
255 : &sse8x8[k], &sum8x8[k]);
256 : break;
257 : }
258 : } else {
259 : vpx_get8x8var(src + src_stride * i + j, src_stride,
260 : ref + ref_stride * i + j, ref_stride, &sse8x8[k],
261 : &sum8x8[k]);
262 : }
263 : #else
264 0 : vpx_get8x8var(src + src_stride * i + j, src_stride,
265 0 : ref + ref_stride * i + j, ref_stride, &sse8x8[k],
266 0 : &sum8x8[k]);
267 : #endif
268 0 : *sse += sse8x8[k];
269 0 : *sum += sum8x8[k];
270 0 : var8x8[k] = sse8x8[k] - (uint32_t)(((int64_t)sum8x8[k] * sum8x8[k]) >> 6);
271 0 : k++;
272 : }
273 : }
274 0 : }
275 :
276 0 : static void calculate_variance(int bw, int bh, TX_SIZE tx_size,
277 : unsigned int *sse_i, int *sum_i,
278 : unsigned int *var_o, unsigned int *sse_o,
279 : int *sum_o) {
280 0 : const BLOCK_SIZE unit_size = txsize_to_bsize[tx_size];
281 0 : const int nw = 1 << (bw - b_width_log2_lookup[unit_size]);
282 0 : const int nh = 1 << (bh - b_height_log2_lookup[unit_size]);
283 0 : int i, j, k = 0;
284 :
285 0 : for (i = 0; i < nh; i += 2) {
286 0 : for (j = 0; j < nw; j += 2) {
287 0 : sse_o[k] = sse_i[i * nw + j] + sse_i[i * nw + j + 1] +
288 0 : sse_i[(i + 1) * nw + j] + sse_i[(i + 1) * nw + j + 1];
289 0 : sum_o[k] = sum_i[i * nw + j] + sum_i[i * nw + j + 1] +
290 0 : sum_i[(i + 1) * nw + j] + sum_i[(i + 1) * nw + j + 1];
291 0 : var_o[k] = sse_o[k] - (uint32_t)(((int64_t)sum_o[k] * sum_o[k]) >>
292 0 : (b_width_log2_lookup[unit_size] +
293 0 : b_height_log2_lookup[unit_size] + 6));
294 0 : k++;
295 : }
296 : }
297 0 : }
298 :
299 0 : static void model_rd_for_sb_y_large(VP9_COMP *cpi, BLOCK_SIZE bsize,
300 : MACROBLOCK *x, MACROBLOCKD *xd,
301 : int *out_rate_sum, int64_t *out_dist_sum,
302 : unsigned int *var_y, unsigned int *sse_y,
303 : int mi_row, int mi_col, int *early_term) {
304 : // Note our transform coeffs are 8 times an orthogonal transform.
305 : // Hence quantizer step is also 8 times. To get effective quantizer
306 : // we need to divide by 8 before sending to modeling function.
307 : unsigned int sse;
308 : int rate;
309 : int64_t dist;
310 0 : struct macroblock_plane *const p = &x->plane[0];
311 0 : struct macroblockd_plane *const pd = &xd->plane[0];
312 0 : const uint32_t dc_quant = pd->dequant[0];
313 0 : const uint32_t ac_quant = pd->dequant[1];
314 0 : const int64_t dc_thr = dc_quant * dc_quant >> 6;
315 0 : const int64_t ac_thr = ac_quant * ac_quant >> 6;
316 : unsigned int var;
317 : int sum;
318 0 : int skip_dc = 0;
319 :
320 0 : const int bw = b_width_log2_lookup[bsize];
321 0 : const int bh = b_height_log2_lookup[bsize];
322 0 : const int num8x8 = 1 << (bw + bh - 2);
323 0 : unsigned int sse8x8[64] = { 0 };
324 0 : int sum8x8[64] = { 0 };
325 0 : unsigned int var8x8[64] = { 0 };
326 : TX_SIZE tx_size;
327 : int i, k;
328 : #if CONFIG_VP9_HIGHBITDEPTH
329 : const vpx_bit_depth_t bd = cpi->common.bit_depth;
330 : #endif
331 : // Calculate variance for whole partition, and also save 8x8 blocks' variance
332 : // to be used in following transform skipping test.
333 0 : block_variance(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
334 : 4 << bw, 4 << bh, &sse, &sum, 8,
335 : #if CONFIG_VP9_HIGHBITDEPTH
336 : cpi->common.use_highbitdepth, bd,
337 : #endif
338 : sse8x8, sum8x8, var8x8);
339 0 : var = sse - (unsigned int)(((int64_t)sum * sum) >> (bw + bh + 4));
340 :
341 0 : *var_y = var;
342 0 : *sse_y = sse;
343 :
344 0 : if (cpi->common.tx_mode == TX_MODE_SELECT) {
345 0 : if (sse > (var << 2))
346 0 : tx_size = VPXMIN(max_txsize_lookup[bsize],
347 : tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
348 : else
349 0 : tx_size = TX_8X8;
350 :
351 0 : if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
352 0 : cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id))
353 0 : tx_size = TX_8X8;
354 0 : else if (tx_size > TX_16X16)
355 0 : tx_size = TX_16X16;
356 : } else {
357 0 : tx_size = VPXMIN(max_txsize_lookup[bsize],
358 : tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
359 : }
360 :
361 0 : assert(tx_size >= TX_8X8);
362 0 : xd->mi[0]->tx_size = tx_size;
363 :
364 : // Evaluate if the partition block is a skippable block in Y plane.
365 : {
366 0 : unsigned int sse16x16[16] = { 0 };
367 0 : int sum16x16[16] = { 0 };
368 0 : unsigned int var16x16[16] = { 0 };
369 0 : const int num16x16 = num8x8 >> 2;
370 :
371 0 : unsigned int sse32x32[4] = { 0 };
372 0 : int sum32x32[4] = { 0 };
373 0 : unsigned int var32x32[4] = { 0 };
374 0 : const int num32x32 = num8x8 >> 4;
375 :
376 0 : int ac_test = 1;
377 0 : int dc_test = 1;
378 0 : const int num = (tx_size == TX_8X8)
379 : ? num8x8
380 0 : : ((tx_size == TX_16X16) ? num16x16 : num32x32);
381 0 : const unsigned int *sse_tx =
382 : (tx_size == TX_8X8) ? sse8x8
383 0 : : ((tx_size == TX_16X16) ? sse16x16 : sse32x32);
384 0 : const unsigned int *var_tx =
385 : (tx_size == TX_8X8) ? var8x8
386 0 : : ((tx_size == TX_16X16) ? var16x16 : var32x32);
387 :
388 : // Calculate variance if tx_size > TX_8X8
389 0 : if (tx_size >= TX_16X16)
390 0 : calculate_variance(bw, bh, TX_8X8, sse8x8, sum8x8, var16x16, sse16x16,
391 : sum16x16);
392 0 : if (tx_size == TX_32X32)
393 0 : calculate_variance(bw, bh, TX_16X16, sse16x16, sum16x16, var32x32,
394 : sse32x32, sum32x32);
395 :
396 : // Skipping test
397 0 : x->skip_txfm[0] = SKIP_TXFM_NONE;
398 0 : for (k = 0; k < num; k++)
399 : // Check if all ac coefficients can be quantized to zero.
400 0 : if (!(var_tx[k] < ac_thr || var == 0)) {
401 0 : ac_test = 0;
402 0 : break;
403 : }
404 :
405 0 : for (k = 0; k < num; k++)
406 : // Check if dc coefficient can be quantized to zero.
407 0 : if (!(sse_tx[k] - var_tx[k] < dc_thr || sse == var)) {
408 0 : dc_test = 0;
409 0 : break;
410 : }
411 :
412 0 : if (ac_test) {
413 0 : x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
414 :
415 0 : if (dc_test) x->skip_txfm[0] = SKIP_TXFM_AC_DC;
416 0 : } else if (dc_test) {
417 0 : skip_dc = 1;
418 : }
419 : }
420 :
421 0 : if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
422 0 : int skip_uv[2] = { 0 };
423 : unsigned int var_uv[2];
424 : unsigned int sse_uv[2];
425 :
426 0 : *out_rate_sum = 0;
427 0 : *out_dist_sum = sse << 4;
428 :
429 : // Transform skipping test in UV planes.
430 0 : for (i = 1; i <= 2; i++) {
431 0 : struct macroblock_plane *const p = &x->plane[i];
432 0 : struct macroblockd_plane *const pd = &xd->plane[i];
433 0 : const TX_SIZE uv_tx_size = get_uv_tx_size(xd->mi[0], pd);
434 0 : const BLOCK_SIZE unit_size = txsize_to_bsize[uv_tx_size];
435 0 : const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, pd);
436 0 : const int uv_bw = b_width_log2_lookup[uv_bsize];
437 0 : const int uv_bh = b_height_log2_lookup[uv_bsize];
438 0 : const int sf = (uv_bw - b_width_log2_lookup[unit_size]) +
439 0 : (uv_bh - b_height_log2_lookup[unit_size]);
440 0 : const uint32_t uv_dc_thr = pd->dequant[0] * pd->dequant[0] >> (6 - sf);
441 0 : const uint32_t uv_ac_thr = pd->dequant[1] * pd->dequant[1] >> (6 - sf);
442 0 : int j = i - 1;
443 :
444 0 : vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, i);
445 0 : var_uv[j] = cpi->fn_ptr[uv_bsize].vf(
446 0 : p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride, &sse_uv[j]);
447 :
448 0 : if ((var_uv[j] < uv_ac_thr || var_uv[j] == 0) &&
449 0 : (sse_uv[j] - var_uv[j] < uv_dc_thr || sse_uv[j] == var_uv[j]))
450 0 : skip_uv[j] = 1;
451 : else
452 : break;
453 : }
454 :
455 : // If the transform in YUV planes are skippable, the mode search checks
456 : // fewer inter modes and doesn't check intra modes.
457 0 : if (skip_uv[0] & skip_uv[1]) {
458 0 : *early_term = 1;
459 : }
460 :
461 0 : return;
462 : }
463 :
464 0 : if (!skip_dc) {
465 : #if CONFIG_VP9_HIGHBITDEPTH
466 : vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
467 : dc_quant >> (xd->bd - 5), &rate, &dist);
468 : #else
469 0 : vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
470 : dc_quant >> 3, &rate, &dist);
471 : #endif // CONFIG_VP9_HIGHBITDEPTH
472 : }
473 :
474 0 : if (!skip_dc) {
475 0 : *out_rate_sum = rate >> 1;
476 0 : *out_dist_sum = dist << 3;
477 : } else {
478 0 : *out_rate_sum = 0;
479 0 : *out_dist_sum = (sse - var) << 4;
480 : }
481 :
482 : #if CONFIG_VP9_HIGHBITDEPTH
483 : vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
484 : ac_quant >> (xd->bd - 5), &rate, &dist);
485 : #else
486 0 : vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], ac_quant >> 3,
487 : &rate, &dist);
488 : #endif // CONFIG_VP9_HIGHBITDEPTH
489 :
490 0 : *out_rate_sum += rate;
491 0 : *out_dist_sum += dist << 4;
492 : }
493 :
494 0 : static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x,
495 : MACROBLOCKD *xd, int *out_rate_sum,
496 : int64_t *out_dist_sum, unsigned int *var_y,
497 : unsigned int *sse_y) {
498 : // Note our transform coeffs are 8 times an orthogonal transform.
499 : // Hence quantizer step is also 8 times. To get effective quantizer
500 : // we need to divide by 8 before sending to modeling function.
501 : unsigned int sse;
502 : int rate;
503 : int64_t dist;
504 0 : struct macroblock_plane *const p = &x->plane[0];
505 0 : struct macroblockd_plane *const pd = &xd->plane[0];
506 0 : const int64_t dc_thr = p->quant_thred[0] >> 6;
507 0 : const int64_t ac_thr = p->quant_thred[1] >> 6;
508 0 : const uint32_t dc_quant = pd->dequant[0];
509 0 : const uint32_t ac_quant = pd->dequant[1];
510 0 : unsigned int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride,
511 0 : pd->dst.buf, pd->dst.stride, &sse);
512 0 : int skip_dc = 0;
513 :
514 0 : *var_y = var;
515 0 : *sse_y = sse;
516 :
517 0 : if (cpi->common.tx_mode == TX_MODE_SELECT) {
518 0 : if (sse > (var << 2))
519 0 : xd->mi[0]->tx_size =
520 0 : VPXMIN(max_txsize_lookup[bsize],
521 : tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
522 : else
523 0 : xd->mi[0]->tx_size = TX_8X8;
524 :
525 0 : if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
526 0 : cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id))
527 0 : xd->mi[0]->tx_size = TX_8X8;
528 0 : else if (xd->mi[0]->tx_size > TX_16X16)
529 0 : xd->mi[0]->tx_size = TX_16X16;
530 : } else {
531 0 : xd->mi[0]->tx_size =
532 0 : VPXMIN(max_txsize_lookup[bsize],
533 : tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
534 : }
535 :
536 : // Evaluate if the partition block is a skippable block in Y plane.
537 : {
538 0 : const BLOCK_SIZE unit_size = txsize_to_bsize[xd->mi[0]->tx_size];
539 0 : const unsigned int num_blk_log2 =
540 0 : (b_width_log2_lookup[bsize] - b_width_log2_lookup[unit_size]) +
541 0 : (b_height_log2_lookup[bsize] - b_height_log2_lookup[unit_size]);
542 0 : const unsigned int sse_tx = sse >> num_blk_log2;
543 0 : const unsigned int var_tx = var >> num_blk_log2;
544 :
545 0 : x->skip_txfm[0] = SKIP_TXFM_NONE;
546 : // Check if all ac coefficients can be quantized to zero.
547 0 : if (var_tx < ac_thr || var == 0) {
548 0 : x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
549 : // Check if dc coefficient can be quantized to zero.
550 0 : if (sse_tx - var_tx < dc_thr || sse == var)
551 0 : x->skip_txfm[0] = SKIP_TXFM_AC_DC;
552 : } else {
553 0 : if (sse_tx - var_tx < dc_thr || sse == var) skip_dc = 1;
554 : }
555 : }
556 :
557 0 : if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
558 0 : *out_rate_sum = 0;
559 0 : *out_dist_sum = sse << 4;
560 0 : return;
561 : }
562 :
563 0 : if (!skip_dc) {
564 : #if CONFIG_VP9_HIGHBITDEPTH
565 : vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
566 : dc_quant >> (xd->bd - 5), &rate, &dist);
567 : #else
568 0 : vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
569 : dc_quant >> 3, &rate, &dist);
570 : #endif // CONFIG_VP9_HIGHBITDEPTH
571 : }
572 :
573 0 : if (!skip_dc) {
574 0 : *out_rate_sum = rate >> 1;
575 0 : *out_dist_sum = dist << 3;
576 : } else {
577 0 : *out_rate_sum = 0;
578 0 : *out_dist_sum = (sse - var) << 4;
579 : }
580 :
581 : #if CONFIG_VP9_HIGHBITDEPTH
582 : vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
583 : ac_quant >> (xd->bd - 5), &rate, &dist);
584 : #else
585 0 : vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], ac_quant >> 3,
586 : &rate, &dist);
587 : #endif // CONFIG_VP9_HIGHBITDEPTH
588 :
589 0 : *out_rate_sum += rate;
590 0 : *out_dist_sum += dist << 4;
591 : }
592 :
593 : #if CONFIG_VP9_HIGHBITDEPTH
594 : static void block_yrd(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *this_rdc,
595 : int *skippable, int64_t *sse, BLOCK_SIZE bsize,
596 : TX_SIZE tx_size) {
597 : MACROBLOCKD *xd = &x->e_mbd;
598 : unsigned int var_y, sse_y;
599 :
600 : (void)tx_size;
601 : model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc->rate, &this_rdc->dist, &var_y,
602 : &sse_y);
603 : *sse = INT_MAX;
604 : *skippable = 0;
605 : return;
606 : }
607 : #else
608 0 : static void block_yrd(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *this_rdc,
609 : int *skippable, int64_t *sse, BLOCK_SIZE bsize,
610 : TX_SIZE tx_size) {
611 0 : MACROBLOCKD *xd = &x->e_mbd;
612 0 : const struct macroblockd_plane *pd = &xd->plane[0];
613 0 : struct macroblock_plane *const p = &x->plane[0];
614 0 : const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
615 0 : const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
616 0 : const int step = 1 << (tx_size << 1);
617 0 : const int block_step = (1 << tx_size);
618 0 : int block = 0, r, c;
619 0 : const int max_blocks_wide =
620 0 : num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 : xd->mb_to_right_edge >> 5);
621 0 : const int max_blocks_high =
622 0 : num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 : xd->mb_to_bottom_edge >> 5);
623 0 : int eob_cost = 0;
624 0 : const int bw = 4 * num_4x4_w;
625 0 : const int bh = 4 * num_4x4_h;
626 :
627 : (void)cpi;
628 :
629 : // The max tx_size passed in is TX_16X16.
630 0 : assert(tx_size != TX_32X32);
631 :
632 0 : vpx_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
633 0 : pd->dst.buf, pd->dst.stride);
634 0 : *skippable = 1;
635 : // Keep track of the row and column of the blocks we use so that we know
636 : // if we are in the unrestricted motion border.
637 0 : for (r = 0; r < max_blocks_high; r += block_step) {
638 0 : for (c = 0; c < num_4x4_w; c += block_step) {
639 0 : if (c < max_blocks_wide) {
640 0 : const scan_order *const scan_order = &vp9_default_scan_orders[tx_size];
641 0 : tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
642 0 : tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
643 0 : tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
644 0 : uint16_t *const eob = &p->eobs[block];
645 0 : const int diff_stride = bw;
646 : const int16_t *src_diff;
647 0 : src_diff = &p->src_diff[(r * diff_stride + c) << 2];
648 :
649 0 : switch (tx_size) {
650 : case TX_16X16:
651 0 : vpx_hadamard_16x16(src_diff, diff_stride, (int16_t *)coeff);
652 0 : vp9_quantize_fp(coeff, 256, x->skip_block, p->zbin, p->round_fp,
653 0 : p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
654 : pd->dequant, eob, scan_order->scan,
655 : scan_order->iscan);
656 0 : break;
657 : case TX_8X8:
658 0 : vpx_hadamard_8x8(src_diff, diff_stride, (int16_t *)coeff);
659 0 : vp9_quantize_fp(coeff, 64, x->skip_block, p->zbin, p->round_fp,
660 0 : p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
661 : pd->dequant, eob, scan_order->scan,
662 : scan_order->iscan);
663 0 : break;
664 : case TX_4X4:
665 0 : x->fwd_txm4x4(src_diff, coeff, diff_stride);
666 0 : vp9_quantize_fp(coeff, 16, x->skip_block, p->zbin, p->round_fp,
667 0 : p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
668 : pd->dequant, eob, scan_order->scan,
669 : scan_order->iscan);
670 0 : break;
671 0 : default: assert(0); break;
672 : }
673 0 : *skippable &= (*eob == 0);
674 0 : eob_cost += 1;
675 : }
676 0 : block += step;
677 : }
678 : }
679 :
680 0 : this_rdc->rate = 0;
681 0 : if (*sse < INT64_MAX) {
682 0 : *sse = (*sse << 6) >> 2;
683 0 : if (*skippable) {
684 0 : this_rdc->dist = *sse;
685 0 : return;
686 : }
687 : }
688 :
689 0 : block = 0;
690 0 : this_rdc->dist = 0;
691 0 : for (r = 0; r < max_blocks_high; r += block_step) {
692 0 : for (c = 0; c < num_4x4_w; c += block_step) {
693 0 : if (c < max_blocks_wide) {
694 0 : tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
695 0 : tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
696 0 : tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
697 0 : uint16_t *const eob = &p->eobs[block];
698 :
699 0 : if (*eob == 1)
700 0 : this_rdc->rate += (int)abs(qcoeff[0]);
701 0 : else if (*eob > 1)
702 0 : this_rdc->rate += vpx_satd((const int16_t *)qcoeff, step << 4);
703 :
704 0 : this_rdc->dist += vp9_block_error_fp(coeff, dqcoeff, step << 4) >> 2;
705 : }
706 0 : block += step;
707 : }
708 : }
709 :
710 : // If skippable is set, rate gets clobbered later.
711 0 : this_rdc->rate <<= (2 + VP9_PROB_COST_SHIFT);
712 0 : this_rdc->rate += (eob_cost << VP9_PROB_COST_SHIFT);
713 : }
714 : #endif
715 :
716 0 : static void model_rd_for_sb_uv(VP9_COMP *cpi, BLOCK_SIZE plane_bsize,
717 : MACROBLOCK *x, MACROBLOCKD *xd,
718 : RD_COST *this_rdc, unsigned int *var_y,
719 : unsigned int *sse_y, int start_plane,
720 : int stop_plane) {
721 : // Note our transform coeffs are 8 times an orthogonal transform.
722 : // Hence quantizer step is also 8 times. To get effective quantizer
723 : // we need to divide by 8 before sending to modeling function.
724 : unsigned int sse;
725 : int rate;
726 : int64_t dist;
727 : int i;
728 : #if CONFIG_VP9_HIGHBITDEPTH
729 : uint64_t tot_var = *var_y;
730 : uint64_t tot_sse = *sse_y;
731 : #else
732 0 : uint32_t tot_var = *var_y;
733 0 : uint32_t tot_sse = *sse_y;
734 : #endif
735 :
736 0 : this_rdc->rate = 0;
737 0 : this_rdc->dist = 0;
738 :
739 0 : for (i = start_plane; i <= stop_plane; ++i) {
740 0 : struct macroblock_plane *const p = &x->plane[i];
741 0 : struct macroblockd_plane *const pd = &xd->plane[i];
742 0 : const uint32_t dc_quant = pd->dequant[0];
743 0 : const uint32_t ac_quant = pd->dequant[1];
744 0 : const BLOCK_SIZE bs = plane_bsize;
745 : unsigned int var;
746 0 : if (!x->color_sensitivity[i - 1]) continue;
747 :
748 0 : var = cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride, pd->dst.buf,
749 : pd->dst.stride, &sse);
750 0 : assert(sse >= var);
751 0 : tot_var += var;
752 0 : tot_sse += sse;
753 :
754 : #if CONFIG_VP9_HIGHBITDEPTH
755 : vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
756 : dc_quant >> (xd->bd - 5), &rate, &dist);
757 : #else
758 0 : vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
759 : dc_quant >> 3, &rate, &dist);
760 : #endif // CONFIG_VP9_HIGHBITDEPTH
761 :
762 0 : this_rdc->rate += rate >> 1;
763 0 : this_rdc->dist += dist << 3;
764 :
765 : #if CONFIG_VP9_HIGHBITDEPTH
766 : vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs],
767 : ac_quant >> (xd->bd - 5), &rate, &dist);
768 : #else
769 0 : vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs], ac_quant >> 3,
770 : &rate, &dist);
771 : #endif // CONFIG_VP9_HIGHBITDEPTH
772 :
773 0 : this_rdc->rate += rate;
774 0 : this_rdc->dist += dist << 4;
775 : }
776 :
777 : #if CONFIG_VP9_HIGHBITDEPTH
778 : *var_y = tot_var > UINT32_MAX ? UINT32_MAX : (uint32_t)tot_var;
779 : *sse_y = tot_sse > UINT32_MAX ? UINT32_MAX : (uint32_t)tot_sse;
780 : #else
781 0 : *var_y = tot_var;
782 0 : *sse_y = tot_sse;
783 : #endif
784 0 : }
785 :
786 0 : static int get_pred_buffer(PRED_BUFFER *p, int len) {
787 : int i;
788 :
789 0 : for (i = 0; i < len; i++) {
790 0 : if (!p[i].in_use) {
791 0 : p[i].in_use = 1;
792 0 : return i;
793 : }
794 : }
795 0 : return -1;
796 : }
797 :
798 0 : static void free_pred_buffer(PRED_BUFFER *p) {
799 0 : if (p != NULL) p->in_use = 0;
800 0 : }
801 :
802 0 : static void encode_breakout_test(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
803 : int mi_row, int mi_col,
804 : MV_REFERENCE_FRAME ref_frame,
805 : PREDICTION_MODE this_mode, unsigned int var_y,
806 : unsigned int sse_y,
807 : struct buf_2d yv12_mb[][MAX_MB_PLANE],
808 : int *rate, int64_t *dist) {
809 0 : MACROBLOCKD *xd = &x->e_mbd;
810 0 : MODE_INFO *const mi = xd->mi[0];
811 0 : const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
812 0 : unsigned int var = var_y, sse = sse_y;
813 : // Skipping threshold for ac.
814 : unsigned int thresh_ac;
815 : // Skipping threshold for dc.
816 : unsigned int thresh_dc;
817 0 : int motion_low = 1;
818 0 : if (mi->mv[0].as_mv.row > 64 || mi->mv[0].as_mv.row < -64 ||
819 0 : mi->mv[0].as_mv.col > 64 || mi->mv[0].as_mv.col < -64)
820 0 : motion_low = 0;
821 0 : if (x->encode_breakout > 0 && motion_low == 1) {
822 : // Set a maximum for threshold to avoid big PSNR loss in low bit rate
823 : // case. Use extreme low threshold for static frames to limit
824 : // skipping.
825 0 : const unsigned int max_thresh = 36000;
826 : // The encode_breakout input
827 0 : const unsigned int min_thresh =
828 0 : VPXMIN(((unsigned int)x->encode_breakout << 4), max_thresh);
829 : #if CONFIG_VP9_HIGHBITDEPTH
830 : const int shift = (xd->bd << 1) - 16;
831 : #endif
832 :
833 : // Calculate threshold according to dequant value.
834 0 : thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) >> 3;
835 : #if CONFIG_VP9_HIGHBITDEPTH
836 : if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
837 : thresh_ac = ROUND_POWER_OF_TWO(thresh_ac, shift);
838 : }
839 : #endif // CONFIG_VP9_HIGHBITDEPTH
840 0 : thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);
841 :
842 : // Adjust ac threshold according to partition size.
843 0 : thresh_ac >>=
844 0 : 8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
845 :
846 0 : thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
847 : #if CONFIG_VP9_HIGHBITDEPTH
848 : if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
849 : thresh_dc = ROUND_POWER_OF_TWO(thresh_dc, shift);
850 : }
851 : #endif // CONFIG_VP9_HIGHBITDEPTH
852 : } else {
853 0 : thresh_ac = 0;
854 0 : thresh_dc = 0;
855 : }
856 :
857 : // Y skipping condition checking for ac and dc.
858 0 : if (var <= thresh_ac && (sse - var) <= thresh_dc) {
859 : unsigned int sse_u, sse_v;
860 : unsigned int var_u, var_v;
861 0 : unsigned int thresh_ac_uv = thresh_ac;
862 0 : unsigned int thresh_dc_uv = thresh_dc;
863 0 : if (x->sb_is_skin) {
864 0 : thresh_ac_uv = 0;
865 0 : thresh_dc_uv = 0;
866 : }
867 :
868 : // Skip UV prediction unless breakout is zero (lossless) to save
869 : // computation with low impact on the result
870 0 : if (x->encode_breakout == 0) {
871 0 : xd->plane[1].pre[0] = yv12_mb[ref_frame][1];
872 0 : xd->plane[2].pre[0] = yv12_mb[ref_frame][2];
873 0 : vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, bsize);
874 : }
875 :
876 0 : var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf, x->plane[1].src.stride,
877 0 : xd->plane[1].dst.buf,
878 : xd->plane[1].dst.stride, &sse_u);
879 :
880 : // U skipping condition checking
881 0 : if (((var_u << 2) <= thresh_ac_uv) && (sse_u - var_u <= thresh_dc_uv)) {
882 0 : var_v = cpi->fn_ptr[uv_size].vf(
883 0 : x->plane[2].src.buf, x->plane[2].src.stride, xd->plane[2].dst.buf,
884 : xd->plane[2].dst.stride, &sse_v);
885 :
886 : // V skipping condition checking
887 0 : if (((var_v << 2) <= thresh_ac_uv) && (sse_v - var_v <= thresh_dc_uv)) {
888 0 : x->skip = 1;
889 :
890 : // The cost of skip bit needs to be added.
891 0 : *rate = cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
892 0 : [INTER_OFFSET(this_mode)];
893 :
894 : // More on this part of rate
895 : // rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
896 :
897 : // Scaling factor for SSE from spatial domain to frequency
898 : // domain is 16. Adjust distortion accordingly.
899 : // TODO(yunqingwang): In this function, only y-plane dist is
900 : // calculated.
901 0 : *dist = (sse << 4); // + ((sse_u + sse_v) << 4);
902 :
903 : // *disable_skip = 1;
904 : }
905 : }
906 : }
907 0 : }
908 :
909 : struct estimate_block_intra_args {
910 : VP9_COMP *cpi;
911 : MACROBLOCK *x;
912 : PREDICTION_MODE mode;
913 : int skippable;
914 : RD_COST *rdc;
915 : };
916 :
917 0 : static void estimate_block_intra(int plane, int block, int row, int col,
918 : BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
919 : void *arg) {
920 0 : struct estimate_block_intra_args *const args = arg;
921 0 : VP9_COMP *const cpi = args->cpi;
922 0 : MACROBLOCK *const x = args->x;
923 0 : MACROBLOCKD *const xd = &x->e_mbd;
924 0 : struct macroblock_plane *const p = &x->plane[0];
925 0 : struct macroblockd_plane *const pd = &xd->plane[0];
926 0 : const BLOCK_SIZE bsize_tx = txsize_to_bsize[tx_size];
927 0 : uint8_t *const src_buf_base = p->src.buf;
928 0 : uint8_t *const dst_buf_base = pd->dst.buf;
929 0 : const int src_stride = p->src.stride;
930 0 : const int dst_stride = pd->dst.stride;
931 : RD_COST this_rdc;
932 :
933 : (void)block;
934 :
935 0 : p->src.buf = &src_buf_base[4 * (row * src_stride + col)];
936 0 : pd->dst.buf = &dst_buf_base[4 * (row * dst_stride + col)];
937 : // Use source buffer as an approximation for the fully reconstructed buffer.
938 0 : vp9_predict_intra_block(xd, b_width_log2_lookup[plane_bsize], tx_size,
939 0 : args->mode, x->skip_encode ? p->src.buf : pd->dst.buf,
940 0 : x->skip_encode ? src_stride : dst_stride, pd->dst.buf,
941 : dst_stride, col, row, plane);
942 :
943 0 : if (plane == 0) {
944 0 : int64_t this_sse = INT64_MAX;
945 : // TODO(jingning): This needs further refactoring.
946 0 : block_yrd(cpi, x, &this_rdc, &args->skippable, &this_sse, bsize_tx,
947 : VPXMIN(tx_size, TX_16X16));
948 : } else {
949 0 : unsigned int var = 0;
950 0 : unsigned int sse = 0;
951 0 : model_rd_for_sb_uv(cpi, plane_bsize, x, xd, &this_rdc, &var, &sse, plane,
952 : plane);
953 : }
954 :
955 0 : p->src.buf = src_buf_base;
956 0 : pd->dst.buf = dst_buf_base;
957 0 : args->rdc->rate += this_rdc.rate;
958 0 : args->rdc->dist += this_rdc.dist;
959 0 : }
960 :
961 : static const THR_MODES mode_idx[MAX_REF_FRAMES - 1][4] = {
962 : { THR_DC, THR_V_PRED, THR_H_PRED, THR_TM },
963 : { THR_NEARESTMV, THR_NEARMV, THR_ZEROMV, THR_NEWMV },
964 : { THR_NEARESTG, THR_NEARG, THR_ZEROG, THR_NEWG },
965 : };
966 :
967 : static const PREDICTION_MODE intra_mode_list[] = { DC_PRED, V_PRED, H_PRED,
968 : TM_PRED };
969 :
970 0 : static int mode_offset(const PREDICTION_MODE mode) {
971 0 : if (mode >= NEARESTMV) {
972 0 : return INTER_OFFSET(mode);
973 : } else {
974 0 : switch (mode) {
975 0 : case DC_PRED: return 0;
976 0 : case V_PRED: return 1;
977 0 : case H_PRED: return 2;
978 0 : case TM_PRED: return 3;
979 0 : default: return -1;
980 : }
981 : }
982 : }
983 :
984 0 : static INLINE void update_thresh_freq_fact(
985 : VP9_COMP *cpi, TileDataEnc *tile_data, int source_variance,
986 : BLOCK_SIZE bsize, MV_REFERENCE_FRAME ref_frame, THR_MODES best_mode_idx,
987 : PREDICTION_MODE mode) {
988 0 : THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
989 0 : int *freq_fact = &tile_data->thresh_freq_fact[bsize][thr_mode_idx];
990 0 : if (thr_mode_idx == best_mode_idx)
991 0 : *freq_fact -= (*freq_fact >> 4);
992 0 : else if (cpi->sf.limit_newmv_early_exit && mode == NEWMV &&
993 0 : ref_frame == LAST_FRAME && source_variance < 5) {
994 0 : *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC, 32);
995 : } else {
996 0 : *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC,
997 : cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
998 : }
999 0 : }
1000 :
1001 0 : void vp9_pick_intra_mode(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
1002 : BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1003 0 : MACROBLOCKD *const xd = &x->e_mbd;
1004 0 : MODE_INFO *const mi = xd->mi[0];
1005 : RD_COST this_rdc, best_rdc;
1006 : PREDICTION_MODE this_mode;
1007 0 : struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
1008 0 : const TX_SIZE intra_tx_size =
1009 0 : VPXMIN(max_txsize_lookup[bsize],
1010 : tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
1011 0 : MODE_INFO *const mic = xd->mi[0];
1012 : int *bmode_costs;
1013 0 : const MODE_INFO *above_mi = xd->above_mi;
1014 0 : const MODE_INFO *left_mi = xd->left_mi;
1015 0 : const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
1016 0 : const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
1017 0 : bmode_costs = cpi->y_mode_costs[A][L];
1018 :
1019 : (void)ctx;
1020 0 : vp9_rd_cost_reset(&best_rdc);
1021 0 : vp9_rd_cost_reset(&this_rdc);
1022 :
1023 0 : mi->ref_frame[0] = INTRA_FRAME;
1024 : // Initialize interp_filter here so we do not have to check for inter block
1025 : // modes in get_pred_context_switchable_interp()
1026 0 : mi->interp_filter = SWITCHABLE_FILTERS;
1027 :
1028 0 : mi->mv[0].as_int = INVALID_MV;
1029 0 : mi->uv_mode = DC_PRED;
1030 0 : memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
1031 :
1032 : // Change the limit of this loop to add other intra prediction
1033 : // mode tests.
1034 0 : for (this_mode = DC_PRED; this_mode <= H_PRED; ++this_mode) {
1035 0 : this_rdc.dist = this_rdc.rate = 0;
1036 0 : args.mode = this_mode;
1037 0 : args.skippable = 1;
1038 0 : args.rdc = &this_rdc;
1039 0 : mi->tx_size = intra_tx_size;
1040 0 : vp9_foreach_transformed_block_in_plane(xd, bsize, 0, estimate_block_intra,
1041 : &args);
1042 0 : if (args.skippable) {
1043 0 : x->skip_txfm[0] = SKIP_TXFM_AC_DC;
1044 0 : this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1);
1045 : } else {
1046 0 : x->skip_txfm[0] = SKIP_TXFM_NONE;
1047 0 : this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0);
1048 : }
1049 0 : this_rdc.rate += bmode_costs[this_mode];
1050 0 : this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
1051 :
1052 0 : if (this_rdc.rdcost < best_rdc.rdcost) {
1053 0 : best_rdc = this_rdc;
1054 0 : mi->mode = this_mode;
1055 : }
1056 : }
1057 :
1058 0 : *rd_cost = best_rdc;
1059 0 : }
1060 :
1061 0 : static void init_ref_frame_cost(VP9_COMMON *const cm, MACROBLOCKD *const xd,
1062 : int ref_frame_cost[MAX_REF_FRAMES]) {
1063 0 : vpx_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
1064 0 : vpx_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
1065 0 : vpx_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
1066 :
1067 0 : ref_frame_cost[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
1068 0 : ref_frame_cost[LAST_FRAME] = ref_frame_cost[GOLDEN_FRAME] =
1069 0 : ref_frame_cost[ALTREF_FRAME] = vp9_cost_bit(intra_inter_p, 1);
1070 :
1071 0 : ref_frame_cost[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
1072 0 : ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1073 0 : ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1074 0 : ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
1075 0 : ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
1076 0 : }
1077 :
1078 : typedef struct {
1079 : MV_REFERENCE_FRAME ref_frame;
1080 : PREDICTION_MODE pred_mode;
1081 : } REF_MODE;
1082 :
1083 : #define RT_INTER_MODES 12
1084 : static const REF_MODE ref_mode_set[RT_INTER_MODES] = {
1085 : { LAST_FRAME, ZEROMV }, { LAST_FRAME, NEARESTMV },
1086 : { GOLDEN_FRAME, ZEROMV }, { LAST_FRAME, NEARMV },
1087 : { LAST_FRAME, NEWMV }, { GOLDEN_FRAME, NEARESTMV },
1088 : { GOLDEN_FRAME, NEARMV }, { GOLDEN_FRAME, NEWMV },
1089 : { ALTREF_FRAME, ZEROMV }, { ALTREF_FRAME, NEARESTMV },
1090 : { ALTREF_FRAME, NEARMV }, { ALTREF_FRAME, NEWMV }
1091 : };
1092 : static const REF_MODE ref_mode_set_svc[RT_INTER_MODES] = {
1093 : { LAST_FRAME, ZEROMV }, { GOLDEN_FRAME, ZEROMV },
1094 : { LAST_FRAME, NEARESTMV }, { LAST_FRAME, NEARMV },
1095 : { GOLDEN_FRAME, NEARESTMV }, { GOLDEN_FRAME, NEARMV },
1096 : { LAST_FRAME, NEWMV }, { GOLDEN_FRAME, NEWMV }
1097 : };
1098 :
1099 0 : static int set_intra_cost_penalty(const VP9_COMP *const cpi, BLOCK_SIZE bsize) {
1100 0 : const VP9_COMMON *const cm = &cpi->common;
1101 : // Reduce the intra cost penalty for small blocks (<=16x16).
1102 0 : int reduction_fac =
1103 0 : (bsize <= BLOCK_16X16) ? ((bsize <= BLOCK_8X8) ? 4 : 2) : 0;
1104 0 : if (cpi->noise_estimate.enabled && cpi->noise_estimate.level == kHigh)
1105 : // Don't reduce intra cost penalty if estimated noise level is high.
1106 0 : reduction_fac = 0;
1107 0 : return vp9_get_intra_cost_penalty(cm->base_qindex, cm->y_dc_delta_q,
1108 0 : cm->bit_depth) >>
1109 : reduction_fac;
1110 : }
1111 :
1112 0 : static INLINE void find_predictors(
1113 : VP9_COMP *cpi, MACROBLOCK *x, MV_REFERENCE_FRAME ref_frame,
1114 : int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1115 : int const_motion[MAX_REF_FRAMES], int *ref_frame_skip_mask,
1116 : const int flag_list[4], TileDataEnc *tile_data, int mi_row, int mi_col,
1117 : struct buf_2d yv12_mb[4][MAX_MB_PLANE], BLOCK_SIZE bsize,
1118 : int force_skip_low_temp_var) {
1119 0 : VP9_COMMON *const cm = &cpi->common;
1120 0 : MACROBLOCKD *const xd = &x->e_mbd;
1121 0 : const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
1122 0 : TileInfo *const tile_info = &tile_data->tile_info;
1123 : // TODO(jingning) placeholder for inter-frame non-RD mode decision.
1124 0 : x->pred_mv_sad[ref_frame] = INT_MAX;
1125 0 : frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
1126 0 : frame_mv[ZEROMV][ref_frame].as_int = 0;
1127 : // this needs various further optimizations. to be continued..
1128 0 : if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
1129 0 : int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
1130 0 : const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
1131 0 : vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
1132 0 : if (cm->use_prev_frame_mvs) {
1133 0 : vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame, candidates, mi_row, mi_col,
1134 0 : x->mbmi_ext->mode_context);
1135 : } else {
1136 0 : const_motion[ref_frame] =
1137 0 : mv_refs_rt(cpi, cm, x, xd, tile_info, xd->mi[0], ref_frame,
1138 0 : candidates, &frame_mv[NEWMV][ref_frame], mi_row, mi_col,
1139 0 : (int)(cpi->svc.use_base_mv && cpi->svc.spatial_layer_id));
1140 : }
1141 0 : vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
1142 0 : &frame_mv[NEARESTMV][ref_frame],
1143 0 : &frame_mv[NEARMV][ref_frame]);
1144 : // Early exit for golden frame if force_skip_low_temp_var is set.
1145 0 : if (!vp9_is_scaled(sf) && bsize >= BLOCK_8X8 &&
1146 0 : !(force_skip_low_temp_var && ref_frame == GOLDEN_FRAME)) {
1147 0 : vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride, ref_frame,
1148 : bsize);
1149 : }
1150 : } else {
1151 0 : *ref_frame_skip_mask |= (1 << ref_frame);
1152 : }
1153 0 : }
1154 :
1155 0 : static void vp9_NEWMV_diff_bias(const NOISE_ESTIMATE *ne, MACROBLOCKD *xd,
1156 : PREDICTION_MODE this_mode, RD_COST *this_rdc,
1157 : BLOCK_SIZE bsize, int mv_row, int mv_col,
1158 : int is_last_frame) {
1159 : // Bias against MVs associated with NEWMV mode that are very different from
1160 : // top/left neighbors.
1161 0 : if (this_mode == NEWMV) {
1162 : int al_mv_average_row;
1163 : int al_mv_average_col;
1164 : int left_row, left_col;
1165 : int row_diff, col_diff;
1166 0 : int above_mv_valid = 0;
1167 0 : int left_mv_valid = 0;
1168 0 : int above_row = 0;
1169 0 : int above_col = 0;
1170 :
1171 0 : if (xd->above_mi) {
1172 0 : above_mv_valid = xd->above_mi->mv[0].as_int != INVALID_MV;
1173 0 : above_row = xd->above_mi->mv[0].as_mv.row;
1174 0 : above_col = xd->above_mi->mv[0].as_mv.col;
1175 : }
1176 0 : if (xd->left_mi) {
1177 0 : left_mv_valid = xd->left_mi->mv[0].as_int != INVALID_MV;
1178 0 : left_row = xd->left_mi->mv[0].as_mv.row;
1179 0 : left_col = xd->left_mi->mv[0].as_mv.col;
1180 : }
1181 0 : if (above_mv_valid && left_mv_valid) {
1182 0 : al_mv_average_row = (above_row + left_row + 1) >> 1;
1183 0 : al_mv_average_col = (above_col + left_col + 1) >> 1;
1184 0 : } else if (above_mv_valid) {
1185 0 : al_mv_average_row = above_row;
1186 0 : al_mv_average_col = above_col;
1187 0 : } else if (left_mv_valid) {
1188 0 : al_mv_average_row = left_row;
1189 0 : al_mv_average_col = left_col;
1190 : } else {
1191 0 : al_mv_average_row = al_mv_average_col = 0;
1192 : }
1193 0 : row_diff = (al_mv_average_row - mv_row);
1194 0 : col_diff = (al_mv_average_col - mv_col);
1195 0 : if (row_diff > 48 || row_diff < -48 || col_diff > 48 || col_diff < -48) {
1196 0 : if (bsize > BLOCK_32X32)
1197 0 : this_rdc->rdcost = this_rdc->rdcost << 1;
1198 : else
1199 0 : this_rdc->rdcost = 3 * this_rdc->rdcost >> 1;
1200 : }
1201 : }
1202 : // If noise estimation is enabled, and estimated level is above threshold,
1203 : // add a bias to LAST reference with small motion, for large blocks.
1204 0 : if (ne->enabled && ne->level >= kMedium && bsize >= BLOCK_32X32 &&
1205 0 : is_last_frame && mv_row < 8 && mv_row > -8 && mv_col < 8 && mv_col > -8) {
1206 0 : this_rdc->rdcost = 7 * this_rdc->rdcost >> 3;
1207 : }
1208 0 : }
1209 :
1210 : #if CONFIG_VP9_TEMPORAL_DENOISING
1211 : static void vp9_pickmode_ctx_den_update(
1212 : VP9_PICKMODE_CTX_DEN *ctx_den, int64_t zero_last_cost_orig,
1213 : int ref_frame_cost[MAX_REF_FRAMES],
1214 : int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], int reuse_inter_pred,
1215 : TX_SIZE best_tx_size, PREDICTION_MODE best_mode,
1216 : MV_REFERENCE_FRAME best_ref_frame, INTERP_FILTER best_pred_filter,
1217 : uint8_t best_mode_skip_txfm) {
1218 : ctx_den->zero_last_cost_orig = zero_last_cost_orig;
1219 : ctx_den->ref_frame_cost = ref_frame_cost;
1220 : ctx_den->frame_mv = frame_mv;
1221 : ctx_den->reuse_inter_pred = reuse_inter_pred;
1222 : ctx_den->best_tx_size = best_tx_size;
1223 : ctx_den->best_mode = best_mode;
1224 : ctx_den->best_ref_frame = best_ref_frame;
1225 : ctx_den->best_pred_filter = best_pred_filter;
1226 : ctx_den->best_mode_skip_txfm = best_mode_skip_txfm;
1227 : }
1228 :
1229 : static void recheck_zeromv_after_denoising(
1230 : VP9_COMP *cpi, MODE_INFO *const mi, MACROBLOCK *x, MACROBLOCKD *const xd,
1231 : VP9_DENOISER_DECISION decision, VP9_PICKMODE_CTX_DEN *ctx_den,
1232 : struct buf_2d yv12_mb[4][MAX_MB_PLANE], RD_COST *best_rdc, BLOCK_SIZE bsize,
1233 : int mi_row, int mi_col) {
1234 : // If INTRA or GOLDEN reference was selected, re-evaluate ZEROMV on
1235 : // denoised result. Only do this under noise conditions, and if rdcost of
1236 : // ZEROMV onoriginal source is not significantly higher than rdcost of best
1237 : // mode.
1238 : if (cpi->noise_estimate.enabled && cpi->noise_estimate.level > kLow &&
1239 : ctx_den->zero_last_cost_orig < (best_rdc->rdcost << 3) &&
1240 : ((ctx_den->best_ref_frame == INTRA_FRAME && decision >= FILTER_BLOCK) ||
1241 : (ctx_den->best_ref_frame == GOLDEN_FRAME &&
1242 : decision == FILTER_ZEROMV_BLOCK))) {
1243 : // Check if we should pick ZEROMV on denoised signal.
1244 : int rate = 0;
1245 : int64_t dist = 0;
1246 : uint32_t var_y = UINT_MAX;
1247 : uint32_t sse_y = UINT_MAX;
1248 : RD_COST this_rdc;
1249 : mi->mode = ZEROMV;
1250 : mi->ref_frame[0] = LAST_FRAME;
1251 : mi->ref_frame[1] = NONE;
1252 : mi->mv[0].as_int = 0;
1253 : mi->interp_filter = EIGHTTAP;
1254 : xd->plane[0].pre[0] = yv12_mb[LAST_FRAME][0];
1255 : vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1256 : model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist, &var_y, &sse_y);
1257 : this_rdc.rate = rate + ctx_den->ref_frame_cost[LAST_FRAME] +
1258 : cpi->inter_mode_cost[x->mbmi_ext->mode_context[LAST_FRAME]]
1259 : [INTER_OFFSET(ZEROMV)];
1260 : this_rdc.dist = dist;
1261 : this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, rate, dist);
1262 : // Don't switch to ZEROMV if the rdcost for ZEROMV on denoised source
1263 : // is higher than best_ref mode (on original source).
1264 : if (this_rdc.rdcost > best_rdc->rdcost) {
1265 : this_rdc = *best_rdc;
1266 : mi->mode = ctx_den->best_mode;
1267 : mi->ref_frame[0] = ctx_den->best_ref_frame;
1268 : mi->interp_filter = ctx_den->best_pred_filter;
1269 : if (ctx_den->best_ref_frame == INTRA_FRAME) {
1270 : mi->mv[0].as_int = INVALID_MV;
1271 : mi->interp_filter = SWITCHABLE_FILTERS;
1272 : } else if (ctx_den->best_ref_frame == GOLDEN_FRAME) {
1273 : mi->mv[0].as_int =
1274 : ctx_den->frame_mv[ctx_den->best_mode][ctx_den->best_ref_frame]
1275 : .as_int;
1276 : if (ctx_den->reuse_inter_pred) {
1277 : xd->plane[0].pre[0] = yv12_mb[GOLDEN_FRAME][0];
1278 : vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1279 : }
1280 : }
1281 : mi->tx_size = ctx_den->best_tx_size;
1282 : x->skip_txfm[0] = ctx_den->best_mode_skip_txfm;
1283 : } else {
1284 : ctx_den->best_ref_frame = LAST_FRAME;
1285 : *best_rdc = this_rdc;
1286 : }
1287 : }
1288 : }
1289 : #endif // CONFIG_VP9_TEMPORAL_DENOISING
1290 :
1291 0 : static INLINE int get_force_skip_low_temp_var(uint8_t *variance_low, int mi_row,
1292 : int mi_col, BLOCK_SIZE bsize) {
1293 0 : const int i = (mi_row & 0x7) >> 1;
1294 0 : const int j = (mi_col & 0x7) >> 1;
1295 0 : int force_skip_low_temp_var = 0;
1296 : // Set force_skip_low_temp_var based on the block size and block offset.
1297 0 : if (bsize == BLOCK_64X64) {
1298 0 : force_skip_low_temp_var = variance_low[0];
1299 0 : } else if (bsize == BLOCK_64X32) {
1300 0 : if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1301 0 : force_skip_low_temp_var = variance_low[1];
1302 0 : } else if (!(mi_col & 0x7) && (mi_row & 0x7)) {
1303 0 : force_skip_low_temp_var = variance_low[2];
1304 : }
1305 0 : } else if (bsize == BLOCK_32X64) {
1306 0 : if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1307 0 : force_skip_low_temp_var = variance_low[3];
1308 0 : } else if ((mi_col & 0x7) && !(mi_row & 0x7)) {
1309 0 : force_skip_low_temp_var = variance_low[4];
1310 : }
1311 0 : } else if (bsize == BLOCK_32X32) {
1312 0 : if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1313 0 : force_skip_low_temp_var = variance_low[5];
1314 0 : } else if ((mi_col & 0x7) && !(mi_row & 0x7)) {
1315 0 : force_skip_low_temp_var = variance_low[6];
1316 0 : } else if (!(mi_col & 0x7) && (mi_row & 0x7)) {
1317 0 : force_skip_low_temp_var = variance_low[7];
1318 0 : } else if ((mi_col & 0x7) && (mi_row & 0x7)) {
1319 0 : force_skip_low_temp_var = variance_low[8];
1320 : }
1321 0 : } else if (bsize == BLOCK_16X16) {
1322 0 : force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]];
1323 0 : } else if (bsize == BLOCK_32X16) {
1324 : // The col shift index for the second 16x16 block.
1325 0 : const int j2 = ((mi_col + 2) & 0x7) >> 1;
1326 : // Only if each 16x16 block inside has low temporal variance.
1327 0 : force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] &&
1328 0 : variance_low[pos_shift_16x16[i][j2]];
1329 0 : } else if (bsize == BLOCK_16X32) {
1330 : // The row shift index for the second 16x16 block.
1331 0 : const int i2 = ((mi_row + 2) & 0x7) >> 1;
1332 0 : force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] &&
1333 0 : variance_low[pos_shift_16x16[i2][j]];
1334 : }
1335 0 : return force_skip_low_temp_var;
1336 : }
1337 :
1338 0 : void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, TileDataEnc *tile_data,
1339 : int mi_row, int mi_col, RD_COST *rd_cost,
1340 : BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1341 0 : VP9_COMMON *const cm = &cpi->common;
1342 0 : SPEED_FEATURES *const sf = &cpi->sf;
1343 0 : const SVC *const svc = &cpi->svc;
1344 0 : MACROBLOCKD *const xd = &x->e_mbd;
1345 0 : MODE_INFO *const mi = xd->mi[0];
1346 0 : struct macroblockd_plane *const pd = &xd->plane[0];
1347 0 : PREDICTION_MODE best_mode = ZEROMV;
1348 0 : MV_REFERENCE_FRAME ref_frame, best_ref_frame = LAST_FRAME;
1349 : MV_REFERENCE_FRAME usable_ref_frame;
1350 0 : TX_SIZE best_tx_size = TX_SIZES;
1351 0 : INTERP_FILTER best_pred_filter = EIGHTTAP;
1352 : int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1353 : struct buf_2d yv12_mb[4][MAX_MB_PLANE];
1354 : static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
1355 : VP9_ALT_FLAG };
1356 : RD_COST this_rdc, best_rdc;
1357 0 : uint8_t skip_txfm = SKIP_TXFM_NONE, best_mode_skip_txfm = SKIP_TXFM_NONE;
1358 : // var_y and sse_y are saved to be used in skipping checking
1359 0 : unsigned int var_y = UINT_MAX;
1360 0 : unsigned int sse_y = UINT_MAX;
1361 0 : const int intra_cost_penalty = set_intra_cost_penalty(cpi, bsize);
1362 0 : int64_t inter_mode_thresh =
1363 0 : RDCOST(x->rdmult, x->rddiv, intra_cost_penalty, 0);
1364 0 : const int *const rd_threshes = cpi->rd.threshes[mi->segment_id][bsize];
1365 0 : const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
1366 : INTERP_FILTER filter_ref;
1367 0 : const int bsl = mi_width_log2_lookup[bsize];
1368 0 : const int pred_filter_search =
1369 0 : cm->interp_filter == SWITCHABLE
1370 0 : ? (((mi_row + mi_col) >> bsl) +
1371 0 : get_chessboard_index(cm->current_video_frame)) &
1372 : 0x1
1373 0 : : 0;
1374 0 : int const_motion[MAX_REF_FRAMES] = { 0 };
1375 0 : const int bh = num_4x4_blocks_high_lookup[bsize] << 2;
1376 0 : const int bw = num_4x4_blocks_wide_lookup[bsize] << 2;
1377 : // For speed 6, the result of interp filter is reused later in actual encoding
1378 : // process.
1379 : // tmp[3] points to dst buffer, and the other 3 point to allocated buffers.
1380 : PRED_BUFFER tmp[4];
1381 : DECLARE_ALIGNED(16, uint8_t, pred_buf[3 * 64 * 64]);
1382 : #if CONFIG_VP9_HIGHBITDEPTH
1383 : DECLARE_ALIGNED(16, uint16_t, pred_buf_16[3 * 64 * 64]);
1384 : #endif
1385 0 : struct buf_2d orig_dst = pd->dst;
1386 0 : PRED_BUFFER *best_pred = NULL;
1387 0 : PRED_BUFFER *this_mode_pred = NULL;
1388 0 : const int pixels_in_block = bh * bw;
1389 0 : int reuse_inter_pred = cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready;
1390 0 : int ref_frame_skip_mask = 0;
1391 : int idx;
1392 0 : int best_pred_sad = INT_MAX;
1393 0 : int best_early_term = 0;
1394 : int ref_frame_cost[MAX_REF_FRAMES];
1395 0 : int svc_force_zero_mode[3] = { 0 };
1396 0 : int perform_intra_pred = 1;
1397 0 : int use_golden_nonzeromv = 1;
1398 0 : int force_skip_low_temp_var = 0;
1399 0 : int skip_ref_find_pred[4] = { 0 };
1400 : #if CONFIG_VP9_TEMPORAL_DENOISING
1401 : VP9_PICKMODE_CTX_DEN ctx_den;
1402 : int64_t zero_last_cost_orig = INT64_MAX;
1403 : #endif
1404 :
1405 0 : init_ref_frame_cost(cm, xd, ref_frame_cost);
1406 :
1407 0 : if (reuse_inter_pred) {
1408 : int i;
1409 0 : for (i = 0; i < 3; i++) {
1410 : #if CONFIG_VP9_HIGHBITDEPTH
1411 : if (cm->use_highbitdepth)
1412 : tmp[i].data = CONVERT_TO_BYTEPTR(&pred_buf_16[pixels_in_block * i]);
1413 : else
1414 : tmp[i].data = &pred_buf[pixels_in_block * i];
1415 : #else
1416 0 : tmp[i].data = &pred_buf[pixels_in_block * i];
1417 : #endif // CONFIG_VP9_HIGHBITDEPTH
1418 0 : tmp[i].stride = bw;
1419 0 : tmp[i].in_use = 0;
1420 : }
1421 0 : tmp[3].data = pd->dst.buf;
1422 0 : tmp[3].stride = pd->dst.stride;
1423 0 : tmp[3].in_use = 0;
1424 : }
1425 :
1426 0 : x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
1427 0 : x->skip = 0;
1428 :
1429 : // Instead of using vp9_get_pred_context_switchable_interp(xd) to assign
1430 : // filter_ref, we use a less strict condition on assigning filter_ref.
1431 : // This is to reduce the probabily of entering the flow of not assigning
1432 : // filter_ref and then skip filter search.
1433 0 : if (xd->above_mi && is_inter_block(xd->above_mi))
1434 0 : filter_ref = xd->above_mi->interp_filter;
1435 0 : else if (xd->left_mi && is_inter_block(xd->left_mi))
1436 0 : filter_ref = xd->left_mi->interp_filter;
1437 : else
1438 0 : filter_ref = cm->interp_filter;
1439 :
1440 : // initialize mode decisions
1441 0 : vp9_rd_cost_reset(&best_rdc);
1442 0 : vp9_rd_cost_reset(rd_cost);
1443 0 : mi->sb_type = bsize;
1444 0 : mi->ref_frame[0] = NONE;
1445 0 : mi->ref_frame[1] = NONE;
1446 :
1447 0 : mi->tx_size =
1448 0 : VPXMIN(max_txsize_lookup[bsize], tx_mode_to_biggest_tx_size[cm->tx_mode]);
1449 :
1450 0 : if (sf->short_circuit_flat_blocks || sf->limit_newmv_early_exit) {
1451 : #if CONFIG_VP9_HIGHBITDEPTH
1452 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
1453 : x->source_variance = vp9_high_get_sby_perpixel_variance(
1454 : cpi, &x->plane[0].src, bsize, xd->bd);
1455 : else
1456 : #endif // CONFIG_VP9_HIGHBITDEPTH
1457 0 : x->source_variance =
1458 0 : vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1459 : }
1460 :
1461 : #if CONFIG_VP9_TEMPORAL_DENOISING
1462 : if (cpi->oxcf.noise_sensitivity > 0 &&
1463 : cpi->denoiser.denoising_level > kDenLowLow) {
1464 : vp9_denoiser_reset_frame_stats(ctx);
1465 : }
1466 : #endif
1467 :
1468 0 : if (cpi->rc.frames_since_golden == 0 && !cpi->use_svc) {
1469 0 : usable_ref_frame = LAST_FRAME;
1470 : } else {
1471 0 : usable_ref_frame = GOLDEN_FRAME;
1472 : }
1473 :
1474 0 : if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) {
1475 0 : if (cpi->rc.alt_ref_gf_group || cpi->rc.is_src_frame_alt_ref)
1476 0 : usable_ref_frame = ALTREF_FRAME;
1477 :
1478 0 : if (cpi->rc.is_src_frame_alt_ref) {
1479 0 : skip_ref_find_pred[LAST_FRAME] = 1;
1480 0 : skip_ref_find_pred[GOLDEN_FRAME] = 1;
1481 : }
1482 : }
1483 :
1484 : // For svc mode, on spatial_layer_id > 0: if the reference has different scale
1485 : // constrain the inter mode to only test zero motion.
1486 0 : if (cpi->use_svc && svc->force_zero_mode_spatial_ref &&
1487 0 : cpi->svc.spatial_layer_id > 0) {
1488 0 : if (cpi->ref_frame_flags & flag_list[LAST_FRAME]) {
1489 0 : struct scale_factors *const sf = &cm->frame_refs[LAST_FRAME - 1].sf;
1490 0 : if (vp9_is_scaled(sf)) svc_force_zero_mode[LAST_FRAME - 1] = 1;
1491 : }
1492 0 : if (cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) {
1493 0 : struct scale_factors *const sf = &cm->frame_refs[GOLDEN_FRAME - 1].sf;
1494 0 : if (vp9_is_scaled(sf)) svc_force_zero_mode[GOLDEN_FRAME - 1] = 1;
1495 : }
1496 : }
1497 :
1498 0 : if (cpi->sf.short_circuit_low_temp_var) {
1499 0 : force_skip_low_temp_var =
1500 0 : get_force_skip_low_temp_var(&x->variance_low[0], mi_row, mi_col, bsize);
1501 : // If force_skip_low_temp_var is set, and for short circuit mode = 1 and 3,
1502 : // skip golden reference.
1503 0 : if ((cpi->sf.short_circuit_low_temp_var == 1 ||
1504 0 : cpi->sf.short_circuit_low_temp_var == 3) &&
1505 : force_skip_low_temp_var) {
1506 0 : usable_ref_frame = LAST_FRAME;
1507 : }
1508 : }
1509 :
1510 0 : if (!((cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) &&
1511 0 : !svc_force_zero_mode[GOLDEN_FRAME - 1] && !force_skip_low_temp_var))
1512 0 : use_golden_nonzeromv = 0;
1513 :
1514 0 : for (ref_frame = LAST_FRAME; ref_frame <= usable_ref_frame; ++ref_frame) {
1515 0 : if (!skip_ref_find_pred[ref_frame]) {
1516 0 : find_predictors(cpi, x, ref_frame, frame_mv, const_motion,
1517 : &ref_frame_skip_mask, flag_list, tile_data, mi_row,
1518 : mi_col, yv12_mb, bsize, force_skip_low_temp_var);
1519 : }
1520 : }
1521 :
1522 0 : for (idx = 0; idx < RT_INTER_MODES; ++idx) {
1523 0 : int rate_mv = 0;
1524 : int mode_rd_thresh;
1525 : int mode_index;
1526 : int i;
1527 : int64_t this_sse;
1528 : int is_skippable;
1529 0 : int this_early_term = 0;
1530 0 : PREDICTION_MODE this_mode = ref_mode_set[idx].pred_mode;
1531 :
1532 0 : ref_frame = ref_mode_set[idx].ref_frame;
1533 :
1534 0 : if (cpi->use_svc) {
1535 0 : this_mode = ref_mode_set_svc[idx].pred_mode;
1536 0 : ref_frame = ref_mode_set_svc[idx].ref_frame;
1537 : }
1538 0 : if (ref_frame > usable_ref_frame) continue;
1539 0 : if (skip_ref_find_pred[ref_frame]) continue;
1540 :
1541 0 : if (sf->short_circuit_flat_blocks && x->source_variance == 0 &&
1542 : this_mode != NEARESTMV) {
1543 0 : continue;
1544 : }
1545 :
1546 0 : if (!(cpi->sf.inter_mode_mask[bsize] & (1 << this_mode))) continue;
1547 :
1548 0 : if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) {
1549 0 : if (cpi->rc.is_src_frame_alt_ref &&
1550 0 : (ref_frame != ALTREF_FRAME ||
1551 0 : frame_mv[this_mode][ref_frame].as_int != 0))
1552 0 : continue;
1553 :
1554 0 : if (cpi->rc.alt_ref_gf_group &&
1555 0 : cpi->rc.frames_since_golden > (cpi->rc.baseline_gf_interval >> 1) &&
1556 0 : ref_frame == GOLDEN_FRAME &&
1557 0 : frame_mv[this_mode][ref_frame].as_int != 0)
1558 0 : continue;
1559 :
1560 0 : if (cpi->rc.alt_ref_gf_group &&
1561 0 : cpi->rc.frames_since_golden < (cpi->rc.baseline_gf_interval >> 1) &&
1562 0 : ref_frame == ALTREF_FRAME &&
1563 0 : frame_mv[this_mode][ref_frame].as_int != 0)
1564 0 : continue;
1565 : }
1566 :
1567 0 : if (!(cpi->ref_frame_flags & flag_list[ref_frame])) continue;
1568 :
1569 0 : if (const_motion[ref_frame] && this_mode == NEARMV) continue;
1570 :
1571 : // Skip non-zeromv mode search for golden frame if force_skip_low_temp_var
1572 : // is set. If nearestmv for golden frame is 0, zeromv mode will be skipped
1573 : // later.
1574 0 : if (force_skip_low_temp_var && ref_frame == GOLDEN_FRAME &&
1575 0 : frame_mv[this_mode][ref_frame].as_int != 0) {
1576 0 : continue;
1577 : }
1578 :
1579 0 : if (cpi->sf.short_circuit_low_temp_var >= 2 && force_skip_low_temp_var &&
1580 0 : ref_frame == LAST_FRAME && this_mode == NEWMV) {
1581 0 : continue;
1582 : }
1583 :
1584 0 : if (cpi->use_svc) {
1585 0 : if (svc_force_zero_mode[ref_frame - 1] &&
1586 0 : frame_mv[this_mode][ref_frame].as_int != 0)
1587 0 : continue;
1588 : }
1589 :
1590 0 : if (sf->reference_masking &&
1591 0 : !(frame_mv[this_mode][ref_frame].as_int == 0 &&
1592 : ref_frame == LAST_FRAME)) {
1593 0 : if (usable_ref_frame < ALTREF_FRAME) {
1594 0 : if (!force_skip_low_temp_var && usable_ref_frame > LAST_FRAME) {
1595 0 : i = (ref_frame == LAST_FRAME) ? GOLDEN_FRAME : LAST_FRAME;
1596 0 : if ((cpi->ref_frame_flags & flag_list[i]))
1597 0 : if (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[i] << 1))
1598 0 : ref_frame_skip_mask |= (1 << ref_frame);
1599 : }
1600 0 : } else if (!cpi->rc.is_src_frame_alt_ref &&
1601 0 : !(frame_mv[this_mode][ref_frame].as_int == 0 &&
1602 : ref_frame == ALTREF_FRAME)) {
1603 0 : int ref1 = (ref_frame == GOLDEN_FRAME) ? LAST_FRAME : GOLDEN_FRAME;
1604 0 : int ref2 = (ref_frame == ALTREF_FRAME) ? LAST_FRAME : ALTREF_FRAME;
1605 0 : if (((cpi->ref_frame_flags & flag_list[ref1]) &&
1606 0 : (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[ref1] << 1))) ||
1607 0 : ((cpi->ref_frame_flags & flag_list[ref2]) &&
1608 0 : (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[ref2] << 1))))
1609 0 : ref_frame_skip_mask |= (1 << ref_frame);
1610 : }
1611 : }
1612 0 : if (ref_frame_skip_mask & (1 << ref_frame)) continue;
1613 :
1614 : // Select prediction reference frames.
1615 0 : for (i = 0; i < MAX_MB_PLANE; i++)
1616 0 : xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
1617 :
1618 0 : mi->ref_frame[0] = ref_frame;
1619 0 : set_ref_ptrs(cm, xd, ref_frame, NONE);
1620 :
1621 0 : mode_index = mode_idx[ref_frame][INTER_OFFSET(this_mode)];
1622 0 : mode_rd_thresh = best_mode_skip_txfm ? rd_threshes[mode_index] << 1
1623 0 : : rd_threshes[mode_index];
1624 :
1625 : // Increase mode_rd_thresh value for GOLDEN_FRAME for improved encoding
1626 : // speed with little/no subjective quality loss.
1627 0 : if (cpi->sf.bias_golden && ref_frame == GOLDEN_FRAME &&
1628 0 : cpi->rc.frames_since_golden > 4)
1629 0 : mode_rd_thresh = mode_rd_thresh << 3;
1630 :
1631 0 : if (rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
1632 0 : rd_thresh_freq_fact[mode_index]))
1633 0 : continue;
1634 :
1635 0 : if (this_mode == NEWMV) {
1636 0 : if (ref_frame > LAST_FRAME && !cpi->use_svc &&
1637 0 : cpi->oxcf.rc_mode == VPX_CBR) {
1638 : int tmp_sad;
1639 : uint32_t dis;
1640 : int cost_list[5];
1641 :
1642 0 : if (bsize < BLOCK_16X16) continue;
1643 :
1644 0 : tmp_sad = vp9_int_pro_motion_estimation(cpi, x, bsize, mi_row, mi_col);
1645 :
1646 0 : if (tmp_sad > x->pred_mv_sad[LAST_FRAME]) continue;
1647 0 : if (tmp_sad + (num_pels_log2_lookup[bsize] << 4) > best_pred_sad)
1648 0 : continue;
1649 :
1650 0 : frame_mv[NEWMV][ref_frame].as_int = mi->mv[0].as_int;
1651 0 : rate_mv = vp9_mv_bit_cost(&frame_mv[NEWMV][ref_frame].as_mv,
1652 0 : &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1653 0 : x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1654 0 : frame_mv[NEWMV][ref_frame].as_mv.row >>= 3;
1655 0 : frame_mv[NEWMV][ref_frame].as_mv.col >>= 3;
1656 :
1657 0 : cpi->find_fractional_mv_step(
1658 0 : x, &frame_mv[NEWMV][ref_frame].as_mv,
1659 0 : &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1660 : cpi->common.allow_high_precision_mv, x->errorperbit,
1661 0 : &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
1662 : cpi->sf.mv.subpel_iters_per_step, cond_cost_list(cpi, cost_list),
1663 0 : x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref_frame], NULL, 0,
1664 : 0);
1665 0 : } else if (svc->use_base_mv && svc->spatial_layer_id) {
1666 0 : if (frame_mv[NEWMV][ref_frame].as_int != INVALID_MV) {
1667 0 : const int pre_stride = xd->plane[0].pre[0].stride;
1668 0 : int base_mv_sad = INT_MAX;
1669 0 : const float base_mv_bias = sf->base_mv_aggressive ? 1.5f : 1.0f;
1670 0 : const uint8_t *const pre_buf =
1671 0 : xd->plane[0].pre[0].buf +
1672 0 : (frame_mv[NEWMV][ref_frame].as_mv.row >> 3) * pre_stride +
1673 0 : (frame_mv[NEWMV][ref_frame].as_mv.col >> 3);
1674 0 : base_mv_sad = cpi->fn_ptr[bsize].sdf(
1675 0 : x->plane[0].src.buf, x->plane[0].src.stride, pre_buf, pre_stride);
1676 :
1677 0 : if (base_mv_sad < (int)(base_mv_bias * x->pred_mv_sad[ref_frame])) {
1678 : // Base layer mv is good.
1679 0 : if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1680 0 : &frame_mv[NEWMV][ref_frame], &rate_mv,
1681 : best_rdc.rdcost, 1)) {
1682 0 : continue;
1683 : }
1684 0 : } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1685 0 : &frame_mv[NEWMV][ref_frame],
1686 : &rate_mv, best_rdc.rdcost, 0)) {
1687 0 : continue;
1688 : }
1689 0 : } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1690 0 : &frame_mv[NEWMV][ref_frame],
1691 : &rate_mv, best_rdc.rdcost, 0)) {
1692 0 : continue;
1693 : }
1694 0 : } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1695 0 : &frame_mv[NEWMV][ref_frame], &rate_mv,
1696 : best_rdc.rdcost, 0)) {
1697 0 : continue;
1698 : }
1699 : }
1700 :
1701 : // If use_golden_nonzeromv is false, NEWMV mode is skipped for golden, no
1702 : // need to compute best_pred_sad which is only used to skip golden NEWMV.
1703 0 : if (use_golden_nonzeromv && this_mode == NEWMV && ref_frame == LAST_FRAME &&
1704 0 : frame_mv[NEWMV][LAST_FRAME].as_int != INVALID_MV) {
1705 0 : const int pre_stride = xd->plane[0].pre[0].stride;
1706 0 : const uint8_t *const pre_buf =
1707 0 : xd->plane[0].pre[0].buf +
1708 0 : (frame_mv[NEWMV][LAST_FRAME].as_mv.row >> 3) * pre_stride +
1709 0 : (frame_mv[NEWMV][LAST_FRAME].as_mv.col >> 3);
1710 0 : best_pred_sad = cpi->fn_ptr[bsize].sdf(
1711 0 : x->plane[0].src.buf, x->plane[0].src.stride, pre_buf, pre_stride);
1712 0 : x->pred_mv_sad[LAST_FRAME] = best_pred_sad;
1713 : }
1714 :
1715 0 : if (this_mode != NEARESTMV &&
1716 0 : frame_mv[this_mode][ref_frame].as_int ==
1717 0 : frame_mv[NEARESTMV][ref_frame].as_int)
1718 0 : continue;
1719 :
1720 0 : mi->mode = this_mode;
1721 0 : mi->mv[0].as_int = frame_mv[this_mode][ref_frame].as_int;
1722 :
1723 : // Search for the best prediction filter type, when the resulting
1724 : // motion vector is at sub-pixel accuracy level for luma component, i.e.,
1725 : // the last three bits are all zeros.
1726 0 : if (reuse_inter_pred) {
1727 0 : if (!this_mode_pred) {
1728 0 : this_mode_pred = &tmp[3];
1729 : } else {
1730 0 : this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
1731 0 : pd->dst.buf = this_mode_pred->data;
1732 0 : pd->dst.stride = bw;
1733 : }
1734 : }
1735 :
1736 0 : if ((this_mode == NEWMV || filter_ref == SWITCHABLE) &&
1737 0 : pred_filter_search &&
1738 0 : (ref_frame == LAST_FRAME ||
1739 0 : (ref_frame == GOLDEN_FRAME &&
1740 0 : (cpi->use_svc || cpi->oxcf.rc_mode == VPX_VBR))) &&
1741 0 : (((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07) != 0)) {
1742 : int pf_rate[3];
1743 : int64_t pf_dist[3];
1744 : unsigned int pf_var[3];
1745 : unsigned int pf_sse[3];
1746 : TX_SIZE pf_tx_size[3];
1747 0 : int64_t best_cost = INT64_MAX;
1748 0 : INTERP_FILTER best_filter = SWITCHABLE, filter;
1749 0 : PRED_BUFFER *current_pred = this_mode_pred;
1750 :
1751 0 : for (filter = EIGHTTAP; filter <= EIGHTTAP_SMOOTH; ++filter) {
1752 : int64_t cost;
1753 0 : mi->interp_filter = filter;
1754 0 : vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1755 0 : model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[filter], &pf_dist[filter],
1756 0 : &pf_var[filter], &pf_sse[filter]);
1757 0 : pf_rate[filter] += vp9_get_switchable_rate(cpi, xd);
1758 0 : cost = RDCOST(x->rdmult, x->rddiv, pf_rate[filter], pf_dist[filter]);
1759 0 : pf_tx_size[filter] = mi->tx_size;
1760 0 : if (cost < best_cost) {
1761 0 : best_filter = filter;
1762 0 : best_cost = cost;
1763 0 : skip_txfm = x->skip_txfm[0];
1764 :
1765 0 : if (reuse_inter_pred) {
1766 0 : if (this_mode_pred != current_pred) {
1767 0 : free_pred_buffer(this_mode_pred);
1768 0 : this_mode_pred = current_pred;
1769 : }
1770 0 : current_pred = &tmp[get_pred_buffer(tmp, 3)];
1771 0 : pd->dst.buf = current_pred->data;
1772 0 : pd->dst.stride = bw;
1773 : }
1774 : }
1775 : }
1776 :
1777 0 : if (reuse_inter_pred && this_mode_pred != current_pred)
1778 0 : free_pred_buffer(current_pred);
1779 :
1780 0 : mi->interp_filter = best_filter;
1781 0 : mi->tx_size = pf_tx_size[best_filter];
1782 0 : this_rdc.rate = pf_rate[best_filter];
1783 0 : this_rdc.dist = pf_dist[best_filter];
1784 0 : var_y = pf_var[best_filter];
1785 0 : sse_y = pf_sse[best_filter];
1786 0 : x->skip_txfm[0] = skip_txfm;
1787 0 : if (reuse_inter_pred) {
1788 0 : pd->dst.buf = this_mode_pred->data;
1789 0 : pd->dst.stride = this_mode_pred->stride;
1790 : }
1791 : } else {
1792 : // TODO(jackychen): the low-bitdepth condition causes a segfault in
1793 : // high-bitdepth builds.
1794 : // https://bugs.chromium.org/p/webm/issues/detail?id=1250
1795 : #if CONFIG_VP9_HIGHBITDEPTH
1796 : const int large_block = bsize > BLOCK_32X32;
1797 : #else
1798 0 : const int large_block =
1799 0 : x->sb_is_skin ? bsize > BLOCK_32X32 : bsize >= BLOCK_32X32;
1800 : #endif
1801 0 : mi->interp_filter = (filter_ref == SWITCHABLE) ? EIGHTTAP : filter_ref;
1802 0 : vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1803 :
1804 : // For large partition blocks, extra testing is done.
1805 0 : if (cpi->oxcf.rc_mode == VPX_CBR && large_block &&
1806 0 : !cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id) &&
1807 0 : cm->base_qindex) {
1808 0 : model_rd_for_sb_y_large(cpi, bsize, x, xd, &this_rdc.rate,
1809 : &this_rdc.dist, &var_y, &sse_y, mi_row, mi_col,
1810 : &this_early_term);
1811 : } else {
1812 0 : model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
1813 : &var_y, &sse_y);
1814 : }
1815 : }
1816 :
1817 0 : if (!this_early_term) {
1818 0 : this_sse = (int64_t)sse_y;
1819 0 : block_yrd(cpi, x, &this_rdc, &is_skippable, &this_sse, bsize,
1820 0 : VPXMIN(mi->tx_size, TX_16X16));
1821 0 : x->skip_txfm[0] = is_skippable;
1822 0 : if (is_skippable) {
1823 0 : this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1824 : } else {
1825 0 : if (RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist) <
1826 0 : RDCOST(x->rdmult, x->rddiv, 0, this_sse)) {
1827 0 : this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
1828 : } else {
1829 0 : this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1830 0 : this_rdc.dist = this_sse;
1831 0 : x->skip_txfm[0] = SKIP_TXFM_AC_DC;
1832 : }
1833 : }
1834 :
1835 0 : if (cm->interp_filter == SWITCHABLE) {
1836 0 : if ((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07)
1837 0 : this_rdc.rate += vp9_get_switchable_rate(cpi, xd);
1838 : }
1839 : } else {
1840 0 : this_rdc.rate += cm->interp_filter == SWITCHABLE
1841 : ? vp9_get_switchable_rate(cpi, xd)
1842 0 : : 0;
1843 0 : this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1844 : }
1845 :
1846 0 : if (x->color_sensitivity[0] || x->color_sensitivity[1]) {
1847 : RD_COST rdc_uv;
1848 0 : const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, &xd->plane[1]);
1849 0 : if (x->color_sensitivity[0])
1850 0 : vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 1);
1851 0 : if (x->color_sensitivity[1])
1852 0 : vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 2);
1853 0 : model_rd_for_sb_uv(cpi, uv_bsize, x, xd, &rdc_uv, &var_y, &sse_y, 1, 2);
1854 0 : this_rdc.rate += rdc_uv.rate;
1855 0 : this_rdc.dist += rdc_uv.dist;
1856 : }
1857 :
1858 0 : this_rdc.rate += rate_mv;
1859 0 : this_rdc.rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
1860 0 : [INTER_OFFSET(this_mode)];
1861 0 : this_rdc.rate += ref_frame_cost[ref_frame];
1862 0 : this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
1863 :
1864 : // Bias against NEWMV that is very different from its neighbors, and bias
1865 : // to small motion-lastref for noisy input.
1866 0 : if (cpi->oxcf.rc_mode == VPX_CBR && cpi->oxcf.speed >= 5 &&
1867 0 : cpi->oxcf.content != VP9E_CONTENT_SCREEN) {
1868 0 : vp9_NEWMV_diff_bias(&cpi->noise_estimate, xd, this_mode, &this_rdc, bsize,
1869 0 : frame_mv[this_mode][ref_frame].as_mv.row,
1870 0 : frame_mv[this_mode][ref_frame].as_mv.col,
1871 : ref_frame == LAST_FRAME);
1872 : }
1873 :
1874 : // Skipping checking: test to see if this block can be reconstructed by
1875 : // prediction only.
1876 0 : if (cpi->allow_encode_breakout) {
1877 0 : encode_breakout_test(cpi, x, bsize, mi_row, mi_col, ref_frame, this_mode,
1878 : var_y, sse_y, yv12_mb, &this_rdc.rate,
1879 : &this_rdc.dist);
1880 0 : if (x->skip) {
1881 0 : this_rdc.rate += rate_mv;
1882 0 : this_rdc.rdcost =
1883 0 : RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
1884 : }
1885 : }
1886 :
1887 : #if CONFIG_VP9_TEMPORAL_DENOISING
1888 : if (cpi->oxcf.noise_sensitivity > 0 &&
1889 : cpi->denoiser.denoising_level > kDenLowLow) {
1890 : vp9_denoiser_update_frame_stats(mi, sse_y, this_mode, ctx);
1891 : // Keep track of zero_last cost.
1892 : if (ref_frame == LAST_FRAME && frame_mv[this_mode][ref_frame].as_int == 0)
1893 : zero_last_cost_orig = this_rdc.rdcost;
1894 : }
1895 : #else
1896 : (void)ctx;
1897 : #endif
1898 :
1899 0 : if (this_rdc.rdcost < best_rdc.rdcost || x->skip) {
1900 0 : best_rdc = this_rdc;
1901 0 : best_mode = this_mode;
1902 0 : best_pred_filter = mi->interp_filter;
1903 0 : best_tx_size = mi->tx_size;
1904 0 : best_ref_frame = ref_frame;
1905 0 : best_mode_skip_txfm = x->skip_txfm[0];
1906 0 : best_early_term = this_early_term;
1907 :
1908 0 : if (reuse_inter_pred) {
1909 0 : free_pred_buffer(best_pred);
1910 0 : best_pred = this_mode_pred;
1911 : }
1912 : } else {
1913 0 : if (reuse_inter_pred) free_pred_buffer(this_mode_pred);
1914 : }
1915 :
1916 0 : if (x->skip) break;
1917 :
1918 : // If early termination flag is 1 and at least 2 modes are checked,
1919 : // the mode search is terminated.
1920 0 : if (best_early_term && idx > 0) {
1921 0 : x->skip = 1;
1922 0 : break;
1923 : }
1924 : }
1925 :
1926 0 : mi->mode = best_mode;
1927 0 : mi->interp_filter = best_pred_filter;
1928 0 : mi->tx_size = best_tx_size;
1929 0 : mi->ref_frame[0] = best_ref_frame;
1930 0 : mi->mv[0].as_int = frame_mv[best_mode][best_ref_frame].as_int;
1931 0 : xd->mi[0]->bmi[0].as_mv[0].as_int = mi->mv[0].as_int;
1932 0 : x->skip_txfm[0] = best_mode_skip_txfm;
1933 :
1934 : // For spatial enhancemanent layer: perform intra prediction only if base
1935 : // layer is chosen as the reference. Always perform intra prediction if
1936 : // LAST is the only reference or is_key_frame is set.
1937 0 : if (cpi->svc.spatial_layer_id) {
1938 0 : perform_intra_pred =
1939 0 : cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame ||
1940 0 : !(cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) ||
1941 0 : (!cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame &&
1942 0 : svc_force_zero_mode[best_ref_frame - 1]);
1943 0 : inter_mode_thresh = (inter_mode_thresh << 1) + inter_mode_thresh;
1944 : }
1945 0 : if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR &&
1946 0 : cpi->rc.is_src_frame_alt_ref)
1947 0 : perform_intra_pred = 0;
1948 : // Perform intra prediction search, if the best SAD is above a certain
1949 : // threshold.
1950 0 : if ((!force_skip_low_temp_var || bsize < BLOCK_32X32) && perform_intra_pred &&
1951 0 : (best_rdc.rdcost == INT64_MAX ||
1952 0 : (!x->skip && best_rdc.rdcost > inter_mode_thresh &&
1953 0 : bsize <= cpi->sf.max_intra_bsize))) {
1954 0 : struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
1955 : int i;
1956 0 : TX_SIZE best_intra_tx_size = TX_SIZES;
1957 0 : TX_SIZE intra_tx_size =
1958 0 : VPXMIN(max_txsize_lookup[bsize],
1959 : tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
1960 0 : if (cpi->oxcf.content != VP9E_CONTENT_SCREEN && intra_tx_size > TX_16X16)
1961 0 : intra_tx_size = TX_16X16;
1962 :
1963 0 : if (reuse_inter_pred && best_pred != NULL) {
1964 0 : if (best_pred->data == orig_dst.buf) {
1965 0 : this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
1966 : #if CONFIG_VP9_HIGHBITDEPTH
1967 : if (cm->use_highbitdepth)
1968 : vpx_highbd_convolve_copy(best_pred->data, best_pred->stride,
1969 : this_mode_pred->data, this_mode_pred->stride,
1970 : NULL, 0, NULL, 0, bw, bh, xd->bd);
1971 : else
1972 : vpx_convolve_copy(best_pred->data, best_pred->stride,
1973 : this_mode_pred->data, this_mode_pred->stride, NULL,
1974 : 0, NULL, 0, bw, bh);
1975 : #else
1976 0 : vpx_convolve_copy(best_pred->data, best_pred->stride,
1977 0 : this_mode_pred->data, this_mode_pred->stride, NULL, 0,
1978 : NULL, 0, bw, bh);
1979 : #endif // CONFIG_VP9_HIGHBITDEPTH
1980 0 : best_pred = this_mode_pred;
1981 : }
1982 : }
1983 0 : pd->dst = orig_dst;
1984 :
1985 0 : for (i = 0; i < 4; ++i) {
1986 0 : const PREDICTION_MODE this_mode = intra_mode_list[i];
1987 0 : THR_MODES mode_index = mode_idx[INTRA_FRAME][mode_offset(this_mode)];
1988 0 : int mode_rd_thresh = rd_threshes[mode_index];
1989 0 : if (sf->short_circuit_flat_blocks && x->source_variance == 0 &&
1990 : this_mode != DC_PRED) {
1991 0 : continue;
1992 : }
1993 :
1994 0 : if (!((1 << this_mode) & cpi->sf.intra_y_mode_bsize_mask[bsize]))
1995 0 : continue;
1996 :
1997 0 : if (rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
1998 0 : rd_thresh_freq_fact[mode_index]))
1999 0 : continue;
2000 :
2001 0 : mi->mode = this_mode;
2002 0 : mi->ref_frame[0] = INTRA_FRAME;
2003 0 : this_rdc.dist = this_rdc.rate = 0;
2004 0 : args.mode = this_mode;
2005 0 : args.skippable = 1;
2006 0 : args.rdc = &this_rdc;
2007 0 : mi->tx_size = intra_tx_size;
2008 0 : vp9_foreach_transformed_block_in_plane(xd, bsize, 0, estimate_block_intra,
2009 : &args);
2010 : // Check skip cost here since skippable is not set for for uv, this
2011 : // mirrors the behavior used by inter
2012 0 : if (args.skippable) {
2013 0 : x->skip_txfm[0] = SKIP_TXFM_AC_DC;
2014 0 : this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1);
2015 : } else {
2016 0 : x->skip_txfm[0] = SKIP_TXFM_NONE;
2017 0 : this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0);
2018 : }
2019 : // Inter and intra RD will mismatch in scale for non-screen content.
2020 0 : if (cpi->oxcf.content == VP9E_CONTENT_SCREEN) {
2021 0 : if (x->color_sensitivity[0])
2022 0 : vp9_foreach_transformed_block_in_plane(xd, bsize, 1,
2023 : estimate_block_intra, &args);
2024 0 : if (x->color_sensitivity[1])
2025 0 : vp9_foreach_transformed_block_in_plane(xd, bsize, 2,
2026 : estimate_block_intra, &args);
2027 : }
2028 0 : this_rdc.rate += cpi->mbmode_cost[this_mode];
2029 0 : this_rdc.rate += ref_frame_cost[INTRA_FRAME];
2030 0 : this_rdc.rate += intra_cost_penalty;
2031 0 : this_rdc.rdcost =
2032 0 : RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2033 :
2034 0 : if (this_rdc.rdcost < best_rdc.rdcost) {
2035 0 : best_rdc = this_rdc;
2036 0 : best_mode = this_mode;
2037 0 : best_intra_tx_size = mi->tx_size;
2038 0 : best_ref_frame = INTRA_FRAME;
2039 0 : mi->uv_mode = this_mode;
2040 0 : mi->mv[0].as_int = INVALID_MV;
2041 0 : best_mode_skip_txfm = x->skip_txfm[0];
2042 : }
2043 : }
2044 :
2045 : // Reset mb_mode_info to the best inter mode.
2046 0 : if (best_ref_frame != INTRA_FRAME) {
2047 0 : mi->tx_size = best_tx_size;
2048 : } else {
2049 0 : mi->tx_size = best_intra_tx_size;
2050 : }
2051 : }
2052 :
2053 0 : pd->dst = orig_dst;
2054 0 : mi->mode = best_mode;
2055 0 : mi->ref_frame[0] = best_ref_frame;
2056 0 : x->skip_txfm[0] = best_mode_skip_txfm;
2057 :
2058 0 : if (!is_inter_block(mi)) {
2059 0 : mi->interp_filter = SWITCHABLE_FILTERS;
2060 : }
2061 :
2062 0 : if (reuse_inter_pred && best_pred != NULL) {
2063 0 : if (best_pred->data != orig_dst.buf && is_inter_mode(mi->mode)) {
2064 : #if CONFIG_VP9_HIGHBITDEPTH
2065 : if (cm->use_highbitdepth)
2066 : vpx_highbd_convolve_copy(best_pred->data, best_pred->stride,
2067 : pd->dst.buf, pd->dst.stride, NULL, 0, NULL, 0,
2068 : bw, bh, xd->bd);
2069 : else
2070 : vpx_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf,
2071 : pd->dst.stride, NULL, 0, NULL, 0, bw, bh);
2072 : #else
2073 0 : vpx_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf,
2074 0 : pd->dst.stride, NULL, 0, NULL, 0, bw, bh);
2075 : #endif // CONFIG_VP9_HIGHBITDEPTH
2076 : }
2077 : }
2078 :
2079 : #if CONFIG_VP9_TEMPORAL_DENOISING
2080 : if (cpi->oxcf.noise_sensitivity > 0 && cpi->resize_pending == 0 &&
2081 : cpi->denoiser.denoising_level > kDenLowLow && cpi->denoiser.reset == 0) {
2082 : VP9_DENOISER_DECISION decision = COPY_BLOCK;
2083 : vp9_pickmode_ctx_den_update(&ctx_den, zero_last_cost_orig, ref_frame_cost,
2084 : frame_mv, reuse_inter_pred, best_tx_size,
2085 : best_mode, best_ref_frame, best_pred_filter,
2086 : best_mode_skip_txfm);
2087 : vp9_denoiser_denoise(cpi, x, mi_row, mi_col, bsize, ctx, &decision);
2088 : recheck_zeromv_after_denoising(cpi, mi, x, xd, decision, &ctx_den, yv12_mb,
2089 : &best_rdc, bsize, mi_row, mi_col);
2090 : best_ref_frame = ctx_den.best_ref_frame;
2091 : }
2092 : #endif
2093 :
2094 0 : if (cpi->sf.adaptive_rd_thresh) {
2095 0 : THR_MODES best_mode_idx = mode_idx[best_ref_frame][mode_offset(mi->mode)];
2096 :
2097 0 : if (best_ref_frame == INTRA_FRAME) {
2098 : // Only consider the modes that are included in the intra_mode_list.
2099 0 : int intra_modes = sizeof(intra_mode_list) / sizeof(PREDICTION_MODE);
2100 : int i;
2101 :
2102 : // TODO(yunqingwang): Check intra mode mask and only update freq_fact
2103 : // for those valid modes.
2104 0 : for (i = 0; i < intra_modes; i++) {
2105 0 : update_thresh_freq_fact(cpi, tile_data, x->source_variance, bsize,
2106 0 : INTRA_FRAME, best_mode_idx, intra_mode_list[i]);
2107 : }
2108 : } else {
2109 0 : for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2110 : PREDICTION_MODE this_mode;
2111 0 : if (best_ref_frame != ref_frame) continue;
2112 0 : for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
2113 0 : update_thresh_freq_fact(cpi, tile_data, x->source_variance, bsize,
2114 : ref_frame, best_mode_idx, this_mode);
2115 : }
2116 : }
2117 : }
2118 : }
2119 :
2120 0 : *rd_cost = best_rdc;
2121 0 : }
2122 :
2123 0 : void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x, int mi_row,
2124 : int mi_col, RD_COST *rd_cost, BLOCK_SIZE bsize,
2125 : PICK_MODE_CONTEXT *ctx) {
2126 0 : VP9_COMMON *const cm = &cpi->common;
2127 0 : SPEED_FEATURES *const sf = &cpi->sf;
2128 0 : MACROBLOCKD *const xd = &x->e_mbd;
2129 0 : MODE_INFO *const mi = xd->mi[0];
2130 0 : MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2131 0 : const struct segmentation *const seg = &cm->seg;
2132 0 : MV_REFERENCE_FRAME ref_frame, second_ref_frame = NONE;
2133 0 : MV_REFERENCE_FRAME best_ref_frame = NONE;
2134 0 : unsigned char segment_id = mi->segment_id;
2135 : struct buf_2d yv12_mb[4][MAX_MB_PLANE];
2136 : static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
2137 : VP9_ALT_FLAG };
2138 0 : int64_t best_rd = INT64_MAX;
2139 : b_mode_info bsi[MAX_REF_FRAMES][4];
2140 0 : int ref_frame_skip_mask = 0;
2141 0 : const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
2142 0 : const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
2143 : int idx, idy;
2144 :
2145 0 : x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
2146 0 : ctx->pred_pixel_ready = 0;
2147 :
2148 0 : for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2149 0 : const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
2150 : int_mv dummy_mv[2];
2151 0 : x->pred_mv_sad[ref_frame] = INT_MAX;
2152 :
2153 0 : if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
2154 0 : int_mv *const candidates = mbmi_ext->ref_mvs[ref_frame];
2155 0 : const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
2156 0 : vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf,
2157 : sf);
2158 0 : vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame, candidates, mi_row, mi_col,
2159 0 : mbmi_ext->mode_context);
2160 :
2161 0 : vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
2162 : &dummy_mv[0], &dummy_mv[1]);
2163 : } else {
2164 0 : ref_frame_skip_mask |= (1 << ref_frame);
2165 : }
2166 : }
2167 :
2168 0 : mi->sb_type = bsize;
2169 0 : mi->tx_size = TX_4X4;
2170 0 : mi->uv_mode = DC_PRED;
2171 0 : mi->ref_frame[0] = LAST_FRAME;
2172 0 : mi->ref_frame[1] = NONE;
2173 0 : mi->interp_filter =
2174 0 : cm->interp_filter == SWITCHABLE ? EIGHTTAP : cm->interp_filter;
2175 :
2176 0 : for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2177 0 : int64_t this_rd = 0;
2178 : int plane;
2179 :
2180 0 : if (ref_frame_skip_mask & (1 << ref_frame)) continue;
2181 :
2182 : #if CONFIG_BETTER_HW_COMPATIBILITY
2183 : if ((bsize == BLOCK_8X4 || bsize == BLOCK_4X8) && ref_frame > INTRA_FRAME &&
2184 : vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
2185 : continue;
2186 : #endif
2187 :
2188 : // TODO(jingning, agrange): Scaling reference frame not supported for
2189 : // sub8x8 blocks. Is this supported now?
2190 0 : if (ref_frame > INTRA_FRAME &&
2191 0 : vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
2192 0 : continue;
2193 :
2194 : // If the segment reference frame feature is enabled....
2195 : // then do nothing if the current ref frame is not allowed..
2196 0 : if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
2197 0 : get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame)
2198 0 : continue;
2199 :
2200 0 : mi->ref_frame[0] = ref_frame;
2201 0 : x->skip = 0;
2202 0 : set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
2203 :
2204 : // Select prediction reference frames.
2205 0 : for (plane = 0; plane < MAX_MB_PLANE; plane++)
2206 0 : xd->plane[plane].pre[0] = yv12_mb[ref_frame][plane];
2207 :
2208 0 : for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
2209 0 : for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
2210 : int_mv b_mv[MB_MODE_COUNT];
2211 0 : int64_t b_best_rd = INT64_MAX;
2212 0 : const int i = idy * 2 + idx;
2213 : PREDICTION_MODE this_mode;
2214 : RD_COST this_rdc;
2215 : unsigned int var_y, sse_y;
2216 :
2217 0 : struct macroblock_plane *p = &x->plane[0];
2218 0 : struct macroblockd_plane *pd = &xd->plane[0];
2219 :
2220 0 : const struct buf_2d orig_src = p->src;
2221 0 : const struct buf_2d orig_dst = pd->dst;
2222 : struct buf_2d orig_pre[2];
2223 0 : memcpy(orig_pre, xd->plane[0].pre, sizeof(orig_pre));
2224 :
2225 : // set buffer pointers for sub8x8 motion search.
2226 0 : p->src.buf =
2227 0 : &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
2228 0 : pd->dst.buf =
2229 0 : &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)];
2230 0 : pd->pre[0].buf =
2231 0 : &pd->pre[0]
2232 0 : .buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->pre[0].stride)];
2233 :
2234 0 : b_mv[ZEROMV].as_int = 0;
2235 0 : b_mv[NEWMV].as_int = INVALID_MV;
2236 0 : vp9_append_sub8x8_mvs_for_idx(cm, xd, i, 0, mi_row, mi_col,
2237 : &b_mv[NEARESTMV], &b_mv[NEARMV],
2238 0 : mbmi_ext->mode_context);
2239 :
2240 0 : for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
2241 0 : int b_rate = 0;
2242 0 : xd->mi[0]->bmi[i].as_mv[0].as_int = b_mv[this_mode].as_int;
2243 :
2244 0 : if (this_mode == NEWMV) {
2245 0 : const int step_param = cpi->sf.mv.fullpel_search_step_param;
2246 : MV mvp_full;
2247 : MV tmp_mv;
2248 : int cost_list[5];
2249 0 : const MvLimits tmp_mv_limits = x->mv_limits;
2250 : uint32_t dummy_dist;
2251 :
2252 0 : if (i == 0) {
2253 0 : mvp_full.row = b_mv[NEARESTMV].as_mv.row >> 3;
2254 0 : mvp_full.col = b_mv[NEARESTMV].as_mv.col >> 3;
2255 : } else {
2256 0 : mvp_full.row = xd->mi[0]->bmi[0].as_mv[0].as_mv.row >> 3;
2257 0 : mvp_full.col = xd->mi[0]->bmi[0].as_mv[0].as_mv.col >> 3;
2258 : }
2259 :
2260 0 : vp9_set_mv_search_range(&x->mv_limits,
2261 0 : &mbmi_ext->ref_mvs[0]->as_mv);
2262 :
2263 0 : vp9_full_pixel_search(cpi, x, bsize, &mvp_full, step_param,
2264 : x->sadperbit4, cond_cost_list(cpi, cost_list),
2265 0 : &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
2266 : &tmp_mv, INT_MAX, 0);
2267 :
2268 0 : x->mv_limits = tmp_mv_limits;
2269 :
2270 : // calculate the bit cost on motion vector
2271 0 : mvp_full.row = tmp_mv.row * 8;
2272 0 : mvp_full.col = tmp_mv.col * 8;
2273 :
2274 0 : b_rate += vp9_mv_bit_cost(
2275 0 : &mvp_full, &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
2276 0 : x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2277 :
2278 0 : b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
2279 0 : [INTER_OFFSET(NEWMV)];
2280 0 : if (RDCOST(x->rdmult, x->rddiv, b_rate, 0) > b_best_rd) continue;
2281 :
2282 0 : cpi->find_fractional_mv_step(
2283 0 : x, &tmp_mv, &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
2284 : cpi->common.allow_high_precision_mv, x->errorperbit,
2285 0 : &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
2286 : cpi->sf.mv.subpel_iters_per_step,
2287 0 : cond_cost_list(cpi, cost_list), x->nmvjointcost, x->mvcost,
2288 0 : &dummy_dist, &x->pred_sse[ref_frame], NULL, 0, 0);
2289 :
2290 0 : xd->mi[0]->bmi[i].as_mv[0].as_mv = tmp_mv;
2291 : } else {
2292 0 : b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
2293 0 : [INTER_OFFSET(this_mode)];
2294 : }
2295 :
2296 : #if CONFIG_VP9_HIGHBITDEPTH
2297 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2298 : vp9_highbd_build_inter_predictor(
2299 : pd->pre[0].buf, pd->pre[0].stride, pd->dst.buf, pd->dst.stride,
2300 : &xd->mi[0]->bmi[i].as_mv[0].as_mv, &xd->block_refs[0]->sf,
2301 : 4 * num_4x4_blocks_wide, 4 * num_4x4_blocks_high, 0,
2302 : vp9_filter_kernels[mi->interp_filter], MV_PRECISION_Q3,
2303 : mi_col * MI_SIZE + 4 * (i & 0x01),
2304 : mi_row * MI_SIZE + 4 * (i >> 1), xd->bd);
2305 : } else {
2306 : #endif
2307 0 : vp9_build_inter_predictor(
2308 0 : pd->pre[0].buf, pd->pre[0].stride, pd->dst.buf, pd->dst.stride,
2309 0 : &xd->mi[0]->bmi[i].as_mv[0].as_mv, &xd->block_refs[0]->sf,
2310 : 4 * num_4x4_blocks_wide, 4 * num_4x4_blocks_high, 0,
2311 0 : vp9_filter_kernels[mi->interp_filter], MV_PRECISION_Q3,
2312 0 : mi_col * MI_SIZE + 4 * (i & 0x01),
2313 0 : mi_row * MI_SIZE + 4 * (i >> 1));
2314 :
2315 : #if CONFIG_VP9_HIGHBITDEPTH
2316 : }
2317 : #endif
2318 :
2319 0 : model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
2320 : &var_y, &sse_y);
2321 :
2322 0 : this_rdc.rate += b_rate;
2323 0 : this_rdc.rdcost =
2324 0 : RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2325 0 : if (this_rdc.rdcost < b_best_rd) {
2326 0 : b_best_rd = this_rdc.rdcost;
2327 0 : bsi[ref_frame][i].as_mode = this_mode;
2328 0 : bsi[ref_frame][i].as_mv[0].as_mv = xd->mi[0]->bmi[i].as_mv[0].as_mv;
2329 : }
2330 : } // mode search
2331 :
2332 : // restore source and prediction buffer pointers.
2333 0 : p->src = orig_src;
2334 0 : pd->pre[0] = orig_pre[0];
2335 0 : pd->dst = orig_dst;
2336 0 : this_rd += b_best_rd;
2337 :
2338 0 : xd->mi[0]->bmi[i] = bsi[ref_frame][i];
2339 0 : if (num_4x4_blocks_wide > 1) xd->mi[0]->bmi[i + 1] = xd->mi[0]->bmi[i];
2340 0 : if (num_4x4_blocks_high > 1) xd->mi[0]->bmi[i + 2] = xd->mi[0]->bmi[i];
2341 : }
2342 : } // loop through sub8x8 blocks
2343 :
2344 0 : if (this_rd < best_rd) {
2345 0 : best_rd = this_rd;
2346 0 : best_ref_frame = ref_frame;
2347 : }
2348 : } // reference frames
2349 :
2350 0 : mi->tx_size = TX_4X4;
2351 0 : mi->ref_frame[0] = best_ref_frame;
2352 0 : for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
2353 0 : for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
2354 0 : const int block = idy * 2 + idx;
2355 0 : xd->mi[0]->bmi[block] = bsi[best_ref_frame][block];
2356 0 : if (num_4x4_blocks_wide > 1)
2357 0 : xd->mi[0]->bmi[block + 1] = bsi[best_ref_frame][block];
2358 0 : if (num_4x4_blocks_high > 1)
2359 0 : xd->mi[0]->bmi[block + 2] = bsi[best_ref_frame][block];
2360 : }
2361 : }
2362 0 : mi->mode = xd->mi[0]->bmi[3].as_mode;
2363 0 : ctx->mic = *(xd->mi[0]);
2364 0 : ctx->mbmi_ext = *x->mbmi_ext;
2365 0 : ctx->skip_txfm[0] = SKIP_TXFM_NONE;
2366 0 : ctx->skip = 0;
2367 : // Dummy assignment for speed -5. No effect in speed -6.
2368 0 : rd_cost->rdcost = best_rd;
2369 0 : }
|