OLD | NEW |
1 /* | 1 /* |
2 * Copyright (c) 2014 The WebM project authors. All Rights Reserved. | 2 * Copyright (c) 2014 The WebM project authors. All Rights Reserved. |
3 * | 3 * |
4 * Use of this source code is governed by a BSD-style license | 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 | 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 | 6 * tree. An additional intellectual property rights grant can be found |
7 * in the file PATENTS. All contributing project authors may | 7 * in the file PATENTS. All contributing project authors may |
8 * be found in the AUTHORS file in the root of the source tree. | 8 * be found in the AUTHORS file in the root of the source tree. |
9 */ | 9 */ |
10 | 10 |
11 #include <assert.h> | 11 #include <assert.h> |
12 #include <limits.h> | 12 #include <limits.h> |
13 #include <math.h> | 13 #include <math.h> |
14 #include <stdio.h> | 14 #include <stdio.h> |
15 | 15 |
16 #include "./vp9_rtcd.h" | 16 #include "./vp9_rtcd.h" |
17 | 17 |
18 #include "vpx_mem/vpx_mem.h" | 18 #include "vpx_mem/vpx_mem.h" |
19 | 19 |
20 #include "vp9/common/vp9_common.h" | 20 #include "vp9/common/vp9_common.h" |
21 #include "vp9/common/vp9_mvref_common.h" | 21 #include "vp9/common/vp9_mvref_common.h" |
22 #include "vp9/common/vp9_reconinter.h" | 22 #include "vp9/common/vp9_reconinter.h" |
23 #include "vp9/common/vp9_reconintra.h" | 23 #include "vp9/common/vp9_reconintra.h" |
24 | 24 |
25 #include "vp9/encoder/vp9_onyx_int.h" | 25 #include "vp9/encoder/vp9_onyx_int.h" |
26 #include "vp9/encoder/vp9_ratectrl.h" | 26 #include "vp9/encoder/vp9_ratectrl.h" |
27 #include "vp9/encoder/vp9_rdopt.h" | 27 #include "vp9/encoder/vp9_rdopt.h" |
28 | 28 |
29 static int full_pixel_motion_search(VP9_COMP *cpi, MACROBLOCK *x, | 29 static void full_pixel_motion_search(VP9_COMP *cpi, MACROBLOCK *x, |
30 const TileInfo *const tile, | 30 const TileInfo *const tile, |
31 BLOCK_SIZE bsize, int mi_row, int mi_col, | 31 BLOCK_SIZE bsize, int mi_row, int mi_col, |
32 int_mv *tmp_mv, int *rate_mv) { | 32 int_mv *tmp_mv, int *rate_mv) { |
33 MACROBLOCKD *xd = &x->e_mbd; | 33 MACROBLOCKD *xd = &x->e_mbd; |
34 MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi; | 34 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi; |
35 struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0}}; | 35 struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0}}; |
36 int bestsme = INT_MAX; | 36 int step_param; |
37 int further_steps, step_param; | |
38 int sadpb = x->sadperbit16; | 37 int sadpb = x->sadperbit16; |
39 MV mvp_full; | 38 MV mvp_full; |
40 int ref = mbmi->ref_frame[0]; | 39 int ref = mbmi->ref_frame[0]; |
41 int_mv ref_mv = mbmi->ref_mvs[ref][0]; | 40 const MV ref_mv = mbmi->ref_mvs[ref][0].as_mv; |
42 int i; | 41 int i; |
43 | 42 |
44 int tmp_col_min = x->mv_col_min; | 43 int tmp_col_min = x->mv_col_min; |
45 int tmp_col_max = x->mv_col_max; | 44 int tmp_col_max = x->mv_col_max; |
46 int tmp_row_min = x->mv_row_min; | 45 int tmp_row_min = x->mv_row_min; |
47 int tmp_row_max = x->mv_row_max; | 46 int tmp_row_max = x->mv_row_max; |
48 | 47 |
49 int buf_offset; | |
50 int stride = xd->plane[0].pre[0].stride; | |
51 | |
52 const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi, | 48 const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi, |
53 ref); | 49 ref); |
54 if (scaled_ref_frame) { | 50 if (scaled_ref_frame) { |
55 int i; | 51 int i; |
56 // Swap out the reference frame for a version that's been scaled to | 52 // Swap out the reference frame for a version that's been scaled to |
57 // match the resolution of the current frame, allowing the existing | 53 // match the resolution of the current frame, allowing the existing |
58 // motion search code to be used without additional modifications. | 54 // motion search code to be used without additional modifications. |
59 for (i = 0; i < MAX_MB_PLANE; i++) | 55 for (i = 0; i < MAX_MB_PLANE; i++) |
60 backup_yv12[i] = xd->plane[i].pre[0]; | 56 backup_yv12[i] = xd->plane[i].pre[0]; |
61 | 57 |
62 setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL); | 58 vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL); |
63 } | 59 } |
64 | 60 |
65 vp9_set_mv_search_range(x, &ref_mv.as_mv); | 61 vp9_set_mv_search_range(x, &ref_mv); |
66 | 62 |
67 // TODO(jingning) exploiting adaptive motion search control in non-RD | 63 // TODO(jingning) exploiting adaptive motion search control in non-RD |
68 // mode decision too. | 64 // mode decision too. |
69 step_param = 6; | 65 step_param = 6; |
70 further_steps = (cpi->sf.max_step_search_steps - 1) - step_param; | |
71 | 66 |
72 for (i = LAST_FRAME; i <= LAST_FRAME && cpi->common.show_frame; ++i) { | 67 for (i = LAST_FRAME; i <= LAST_FRAME && cpi->common.show_frame; ++i) { |
73 if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) { | 68 if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) { |
74 tmp_mv->as_int = INVALID_MV; | 69 tmp_mv->as_int = INVALID_MV; |
75 | 70 |
76 if (scaled_ref_frame) { | 71 if (scaled_ref_frame) { |
77 int i; | 72 int i; |
78 for (i = 0; i < MAX_MB_PLANE; i++) | 73 for (i = 0; i < MAX_MB_PLANE; i++) |
79 xd->plane[i].pre[0] = backup_yv12[i]; | 74 xd->plane[i].pre[0] = backup_yv12[i]; |
80 } | 75 } |
81 return INT_MAX; | 76 return; |
82 } | 77 } |
83 } | 78 } |
84 | 79 assert(x->mv_best_ref_index[ref] <= 2); |
85 mvp_full = mbmi->ref_mvs[ref][x->mv_best_ref_index[ref]].as_mv; | 80 if (x->mv_best_ref_index[ref] < 2) |
| 81 mvp_full = mbmi->ref_mvs[ref][x->mv_best_ref_index[ref]].as_mv; |
| 82 else |
| 83 mvp_full = x->pred_mv[ref].as_mv; |
86 | 84 |
87 mvp_full.col >>= 3; | 85 mvp_full.col >>= 3; |
88 mvp_full.row >>= 3; | 86 mvp_full.row >>= 3; |
89 | 87 |
90 if (cpi->sf.search_method == FAST_HEX) { | 88 if (cpi->sf.search_method == FAST_DIAMOND) { |
91 bestsme = vp9_fast_hex_search(x, &mvp_full, step_param, sadpb, | 89 // NOTE: this returns SAD |
92 &cpi->fn_ptr[bsize], 1, | 90 vp9_fast_dia_search(x, &mvp_full, step_param, sadpb, 0, |
93 &ref_mv.as_mv, &tmp_mv->as_mv); | 91 &cpi->fn_ptr[bsize], 1, |
| 92 &ref_mv, &tmp_mv->as_mv); |
| 93 } else if (cpi->sf.search_method == FAST_HEX) { |
| 94 // NOTE: this returns SAD |
| 95 vp9_fast_hex_search(x, &mvp_full, step_param, sadpb, 0, |
| 96 &cpi->fn_ptr[bsize], 1, |
| 97 &ref_mv, &tmp_mv->as_mv); |
94 } else if (cpi->sf.search_method == HEX) { | 98 } else if (cpi->sf.search_method == HEX) { |
95 bestsme = vp9_hex_search(x, &mvp_full, step_param, sadpb, 1, | 99 // NOTE: this returns SAD |
96 &cpi->fn_ptr[bsize], 1, | 100 vp9_hex_search(x, &mvp_full, step_param, sadpb, 1, |
97 &ref_mv.as_mv, &tmp_mv->as_mv); | 101 &cpi->fn_ptr[bsize], 1, |
| 102 &ref_mv, &tmp_mv->as_mv); |
98 } else if (cpi->sf.search_method == SQUARE) { | 103 } else if (cpi->sf.search_method == SQUARE) { |
99 bestsme = vp9_square_search(x, &mvp_full, step_param, sadpb, 1, | 104 // NOTE: this returns SAD |
100 &cpi->fn_ptr[bsize], 1, | 105 vp9_square_search(x, &mvp_full, step_param, sadpb, 1, |
101 &ref_mv.as_mv, &tmp_mv->as_mv); | 106 &cpi->fn_ptr[bsize], 1, |
| 107 &ref_mv, &tmp_mv->as_mv); |
102 } else if (cpi->sf.search_method == BIGDIA) { | 108 } else if (cpi->sf.search_method == BIGDIA) { |
103 bestsme = vp9_bigdia_search(x, &mvp_full, step_param, sadpb, 1, | 109 // NOTE: this returns SAD |
104 &cpi->fn_ptr[bsize], 1, | 110 vp9_bigdia_search(x, &mvp_full, step_param, sadpb, 1, |
105 &ref_mv.as_mv, &tmp_mv->as_mv); | 111 &cpi->fn_ptr[bsize], 1, |
| 112 &ref_mv, &tmp_mv->as_mv); |
106 } else { | 113 } else { |
107 bestsme = vp9_full_pixel_diamond(cpi, x, &mvp_full, step_param, | 114 int further_steps = (cpi->sf.max_step_search_steps - 1) - step_param; |
108 sadpb, further_steps, 1, | 115 // NOTE: this returns variance |
109 &cpi->fn_ptr[bsize], | 116 vp9_full_pixel_diamond(cpi, x, &mvp_full, step_param, |
110 &ref_mv.as_mv, &tmp_mv->as_mv); | 117 sadpb, further_steps, 1, |
| 118 &cpi->fn_ptr[bsize], |
| 119 &ref_mv, &tmp_mv->as_mv); |
111 } | 120 } |
112 x->mv_col_min = tmp_col_min; | 121 x->mv_col_min = tmp_col_min; |
113 x->mv_col_max = tmp_col_max; | 122 x->mv_col_max = tmp_col_max; |
114 x->mv_row_min = tmp_row_min; | 123 x->mv_row_min = tmp_row_min; |
115 x->mv_row_max = tmp_row_max; | 124 x->mv_row_max = tmp_row_max; |
116 | 125 |
117 if (scaled_ref_frame) { | 126 if (scaled_ref_frame) { |
118 int i; | 127 int i; |
119 for (i = 0; i < MAX_MB_PLANE; i++) | 128 for (i = 0; i < MAX_MB_PLANE; i++) |
120 xd->plane[i].pre[0] = backup_yv12[i]; | 129 xd->plane[i].pre[0] = backup_yv12[i]; |
121 } | 130 } |
122 | 131 |
123 // TODO(jingning) This step can be merged into full pixel search step in the | |
124 // re-designed log-diamond search | |
125 buf_offset = tmp_mv->as_mv.row * stride + tmp_mv->as_mv.col; | |
126 | |
127 // Find sad for current vector. | |
128 bestsme = cpi->fn_ptr[bsize].sdf(x->plane[0].src.buf, x->plane[0].src.stride, | |
129 xd->plane[0].pre[0].buf + buf_offset, | |
130 stride, 0x7fffffff); | |
131 | |
132 // scale to 1/8 pixel resolution | |
133 tmp_mv->as_mv.row = tmp_mv->as_mv.row * 8; | |
134 tmp_mv->as_mv.col = tmp_mv->as_mv.col * 8; | |
135 | |
136 // calculate the bit cost on motion vector | 132 // calculate the bit cost on motion vector |
137 *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv.as_mv, | 133 mvp_full.row = tmp_mv->as_mv.row * 8; |
| 134 mvp_full.col = tmp_mv->as_mv.col * 8; |
| 135 *rate_mv = vp9_mv_bit_cost(&mvp_full, &ref_mv, |
138 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT); | 136 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT); |
139 return bestsme; | |
140 } | 137 } |
141 | 138 |
142 static void sub_pixel_motion_search(VP9_COMP *cpi, MACROBLOCK *x, | 139 static void sub_pixel_motion_search(VP9_COMP *cpi, MACROBLOCK *x, |
143 const TileInfo *const tile, | 140 const TileInfo *const tile, |
144 BLOCK_SIZE bsize, int mi_row, int mi_col, | 141 BLOCK_SIZE bsize, int mi_row, int mi_col, |
145 MV *tmp_mv) { | 142 MV *tmp_mv) { |
146 MACROBLOCKD *xd = &x->e_mbd; | 143 MACROBLOCKD *xd = &x->e_mbd; |
147 MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi; | 144 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi; |
148 struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0}}; | 145 struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0}}; |
149 int ref = mbmi->ref_frame[0]; | 146 int ref = mbmi->ref_frame[0]; |
150 MV ref_mv = mbmi->ref_mvs[ref][0].as_mv; | 147 MV ref_mv = mbmi->ref_mvs[ref][0].as_mv; |
151 int dis; | 148 int dis; |
152 | 149 |
153 const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi, | 150 const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi, |
154 ref); | 151 ref); |
155 if (scaled_ref_frame) { | 152 if (scaled_ref_frame) { |
156 int i; | 153 int i; |
157 // Swap out the reference frame for a version that's been scaled to | 154 // Swap out the reference frame for a version that's been scaled to |
158 // match the resolution of the current frame, allowing the existing | 155 // match the resolution of the current frame, allowing the existing |
159 // motion search code to be used without additional modifications. | 156 // motion search code to be used without additional modifications. |
160 for (i = 0; i < MAX_MB_PLANE; i++) | 157 for (i = 0; i < MAX_MB_PLANE; i++) |
161 backup_yv12[i] = xd->plane[i].pre[0]; | 158 backup_yv12[i] = xd->plane[i].pre[0]; |
162 | 159 |
163 setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL); | 160 vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL); |
164 } | 161 } |
165 | 162 |
166 tmp_mv->col >>= 3; | |
167 tmp_mv->row >>= 3; | |
168 | |
169 cpi->find_fractional_mv_step(x, tmp_mv, &ref_mv, | 163 cpi->find_fractional_mv_step(x, tmp_mv, &ref_mv, |
170 cpi->common.allow_high_precision_mv, | 164 cpi->common.allow_high_precision_mv, |
171 x->errorperbit, | 165 x->errorperbit, |
172 &cpi->fn_ptr[bsize], | 166 &cpi->fn_ptr[bsize], |
173 cpi->sf.subpel_force_stop, | 167 cpi->sf.subpel_force_stop, |
174 cpi->sf.subpel_iters_per_step, | 168 cpi->sf.subpel_iters_per_step, |
175 x->nmvjointcost, x->mvcost, | 169 x->nmvjointcost, x->mvcost, |
176 &dis, &x->pred_sse[ref]); | 170 &dis, &x->pred_sse[ref]); |
177 | 171 |
178 if (scaled_ref_frame) { | 172 if (scaled_ref_frame) { |
179 int i; | 173 int i; |
180 for (i = 0; i < MAX_MB_PLANE; i++) | 174 for (i = 0; i < MAX_MB_PLANE; i++) |
181 xd->plane[i].pre[0] = backup_yv12[i]; | 175 xd->plane[i].pre[0] = backup_yv12[i]; |
182 } | 176 } |
| 177 |
| 178 x->pred_mv[ref].as_mv = *tmp_mv; |
| 179 } |
| 180 |
| 181 static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize, |
| 182 MACROBLOCK *x, MACROBLOCKD *xd, |
| 183 int *out_rate_sum, int64_t *out_dist_sum) { |
| 184 // Note our transform coeffs are 8 times an orthogonal transform. |
| 185 // Hence quantizer step is also 8 times. To get effective quantizer |
| 186 // we need to divide by 8 before sending to modeling function. |
| 187 unsigned int sse; |
| 188 int rate; |
| 189 int64_t dist; |
| 190 |
| 191 struct macroblock_plane *const p = &x->plane[0]; |
| 192 struct macroblockd_plane *const pd = &xd->plane[0]; |
| 193 |
| 194 int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride, |
| 195 pd->dst.buf, pd->dst.stride, &sse); |
| 196 |
| 197 vp9_model_rd_from_var_lapndz(sse + var, 1 << num_pels_log2_lookup[bsize], |
| 198 pd->dequant[1] >> 3, &rate, &dist); |
| 199 *out_rate_sum = rate; |
| 200 *out_dist_sum = dist << 3; |
183 } | 201 } |
184 | 202 |
185 // TODO(jingning) placeholder for inter-frame non-RD mode decision. | 203 // TODO(jingning) placeholder for inter-frame non-RD mode decision. |
186 // this needs various further optimizations. to be continued.. | 204 // this needs various further optimizations. to be continued.. |
187 int64_t vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, | 205 int64_t vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, |
188 const TileInfo *const tile, | 206 const TileInfo *const tile, |
189 int mi_row, int mi_col, | 207 int mi_row, int mi_col, |
190 int *returnrate, | 208 int *returnrate, |
191 int64_t *returndistortion, | 209 int64_t *returndistortion, |
192 BLOCK_SIZE bsize) { | 210 BLOCK_SIZE bsize) { |
193 MACROBLOCKD *xd = &x->e_mbd; | 211 MACROBLOCKD *xd = &x->e_mbd; |
194 MB_MODE_INFO *mbmi = &xd->mi_8x8[0]->mbmi; | 212 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi; |
195 struct macroblock_plane *const p = &x->plane[0]; | 213 struct macroblock_plane *const p = &x->plane[0]; |
196 struct macroblockd_plane *const pd = &xd->plane[0]; | 214 struct macroblockd_plane *const pd = &xd->plane[0]; |
197 const BLOCK_SIZE block_size = get_plane_block_size(bsize, &xd->plane[0]); | |
198 MB_PREDICTION_MODE this_mode, best_mode = ZEROMV; | 215 MB_PREDICTION_MODE this_mode, best_mode = ZEROMV; |
199 MV_REFERENCE_FRAME ref_frame, best_ref_frame = LAST_FRAME; | 216 MV_REFERENCE_FRAME ref_frame, best_ref_frame = LAST_FRAME; |
| 217 INTERP_FILTER best_pred_filter = EIGHTTAP; |
200 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES]; | 218 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES]; |
201 struct buf_2d yv12_mb[4][MAX_MB_PLANE]; | 219 struct buf_2d yv12_mb[4][MAX_MB_PLANE]; |
202 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG, | 220 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG, |
203 VP9_ALT_FLAG }; | 221 VP9_ALT_FLAG }; |
204 int64_t best_rd = INT64_MAX; | 222 int64_t best_rd = INT64_MAX; |
205 int64_t this_rd = INT64_MAX; | 223 int64_t this_rd = INT64_MAX; |
206 static const int cost[4]= { 0, 2, 4, 6 }; | |
207 | 224 |
208 const int64_t inter_mode_thresh = 300; | 225 int rate = INT_MAX; |
| 226 int64_t dist = INT64_MAX; |
| 227 |
| 228 VP9_COMMON *cm = &cpi->common; |
| 229 int intra_cost_penalty = 20 * vp9_dc_quant(cm->base_qindex, cm->y_dc_delta_q); |
| 230 |
| 231 const int64_t inter_mode_thresh = RDCOST(x->rdmult, x->rddiv, |
| 232 intra_cost_penalty, 0); |
209 const int64_t intra_mode_cost = 50; | 233 const int64_t intra_mode_cost = 50; |
210 | 234 |
| 235 unsigned char segment_id = mbmi->segment_id; |
| 236 const int *const rd_threshes = cpi->rd_threshes[segment_id][bsize]; |
| 237 const int *const rd_thresh_freq_fact = cpi->rd_thresh_freq_fact[bsize]; |
| 238 // Mode index conversion form THR_MODES to MB_PREDICTION_MODE for a ref frame. |
| 239 int mode_idx[MB_MODE_COUNT] = {0}; |
| 240 INTERP_FILTER filter_ref = SWITCHABLE; |
| 241 |
211 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH; | 242 x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH; |
212 | 243 |
213 x->skip = 0; | 244 x->skip = 0; |
214 if (cpi->active_map_enabled && x->active_ptr[0] == 0) | 245 if (!x->in_active_map) |
215 x->skip = 1; | 246 x->skip = 1; |
216 | |
217 // initialize mode decisions | 247 // initialize mode decisions |
218 *returnrate = INT_MAX; | 248 *returnrate = INT_MAX; |
219 *returndistortion = INT64_MAX; | 249 *returndistortion = INT64_MAX; |
220 vpx_memset(mbmi, 0, sizeof(MB_MODE_INFO)); | 250 vpx_memset(mbmi, 0, sizeof(MB_MODE_INFO)); |
221 mbmi->sb_type = bsize; | 251 mbmi->sb_type = bsize; |
222 mbmi->ref_frame[0] = NONE; | 252 mbmi->ref_frame[0] = NONE; |
223 mbmi->ref_frame[1] = NONE; | 253 mbmi->ref_frame[1] = NONE; |
224 mbmi->tx_size = MIN(max_txsize_lookup[bsize], | 254 mbmi->tx_size = MIN(max_txsize_lookup[bsize], |
225 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]); | 255 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]); |
226 mbmi->interp_filter = cpi->common.interp_filter == SWITCHABLE ? | 256 mbmi->interp_filter = cpi->common.interp_filter == SWITCHABLE ? |
227 EIGHTTAP : cpi->common.interp_filter; | 257 EIGHTTAP : cpi->common.interp_filter; |
228 mbmi->skip = 0; | 258 mbmi->skip = 0; |
229 mbmi->segment_id = 0; | 259 mbmi->segment_id = segment_id; |
230 | 260 |
231 for (ref_frame = LAST_FRAME; ref_frame <= LAST_FRAME ; ++ref_frame) { | 261 for (ref_frame = LAST_FRAME; ref_frame <= LAST_FRAME ; ++ref_frame) { |
232 x->pred_mv_sad[ref_frame] = INT_MAX; | 262 x->pred_mv_sad[ref_frame] = INT_MAX; |
233 if (cpi->ref_frame_flags & flag_list[ref_frame]) { | 263 if (cpi->ref_frame_flags & flag_list[ref_frame]) { |
234 vp9_setup_buffer_inter(cpi, x, tile, | 264 vp9_setup_buffer_inter(cpi, x, tile, |
235 ref_frame, block_size, mi_row, mi_col, | 265 ref_frame, bsize, mi_row, mi_col, |
236 frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb); | 266 frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb); |
237 } | 267 } |
238 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV; | 268 frame_mv[NEWMV][ref_frame].as_int = INVALID_MV; |
239 frame_mv[ZEROMV][ref_frame].as_int = 0; | 269 frame_mv[ZEROMV][ref_frame].as_int = 0; |
240 } | 270 } |
241 | 271 |
| 272 if (xd->up_available) |
| 273 filter_ref = xd->mi[-xd->mi_stride]->mbmi.interp_filter; |
| 274 else if (xd->left_available) |
| 275 filter_ref = xd->mi[-1]->mbmi.interp_filter; |
| 276 |
242 for (ref_frame = LAST_FRAME; ref_frame <= LAST_FRAME ; ++ref_frame) { | 277 for (ref_frame = LAST_FRAME; ref_frame <= LAST_FRAME ; ++ref_frame) { |
243 int rate_mv = 0; | |
244 if (!(cpi->ref_frame_flags & flag_list[ref_frame])) | 278 if (!(cpi->ref_frame_flags & flag_list[ref_frame])) |
245 continue; | 279 continue; |
246 | 280 |
247 // Select prediction reference frames. | 281 // Select prediction reference frames. |
248 xd->plane[0].pre[0] = yv12_mb[ref_frame][0]; | 282 xd->plane[0].pre[0] = yv12_mb[ref_frame][0]; |
249 | 283 |
250 clamp_mv2(&frame_mv[NEARESTMV][ref_frame].as_mv, xd); | 284 clamp_mv2(&frame_mv[NEARESTMV][ref_frame].as_mv, xd); |
251 clamp_mv2(&frame_mv[NEARMV][ref_frame].as_mv, xd); | 285 clamp_mv2(&frame_mv[NEARMV][ref_frame].as_mv, xd); |
252 | 286 |
253 mbmi->ref_frame[0] = ref_frame; | 287 mbmi->ref_frame[0] = ref_frame; |
254 | 288 |
| 289 // Set conversion index for LAST_FRAME. |
| 290 if (ref_frame == LAST_FRAME) { |
| 291 mode_idx[NEARESTMV] = THR_NEARESTMV; // LAST_FRAME, NEARESTMV |
| 292 mode_idx[NEARMV] = THR_NEARMV; // LAST_FRAME, NEARMV |
| 293 mode_idx[ZEROMV] = THR_ZEROMV; // LAST_FRAME, ZEROMV |
| 294 mode_idx[NEWMV] = THR_NEWMV; // LAST_FRAME, NEWMV |
| 295 } |
| 296 |
255 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) { | 297 for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) { |
256 int rate = cost[INTER_OFFSET(this_mode)] | 298 int rate_mv = 0; |
257 << (num_pels_log2_lookup[bsize] - 4); | 299 |
258 int64_t dist; | |
259 if (cpi->sf.disable_inter_mode_mask[bsize] & | 300 if (cpi->sf.disable_inter_mode_mask[bsize] & |
260 (1 << INTER_OFFSET(this_mode))) | 301 (1 << INTER_OFFSET(this_mode))) |
261 continue; | 302 continue; |
262 | 303 |
| 304 if (best_rd < ((int64_t)rd_threshes[mode_idx[this_mode]] * |
| 305 rd_thresh_freq_fact[this_mode] >> 5) || |
| 306 rd_threshes[mode_idx[this_mode]] == INT_MAX) |
| 307 continue; |
| 308 |
263 if (this_mode == NEWMV) { | 309 if (this_mode == NEWMV) { |
| 310 int rate_mode = 0; |
264 if (this_rd < (int64_t)(1 << num_pels_log2_lookup[bsize])) | 311 if (this_rd < (int64_t)(1 << num_pels_log2_lookup[bsize])) |
265 continue; | 312 continue; |
266 | 313 |
267 x->mode_sad[ref_frame][INTER_OFFSET(NEWMV)] = | 314 full_pixel_motion_search(cpi, x, tile, bsize, mi_row, mi_col, |
268 full_pixel_motion_search(cpi, x, tile, bsize, mi_row, mi_col, | 315 &frame_mv[NEWMV][ref_frame], &rate_mv); |
269 &frame_mv[NEWMV][ref_frame], &rate_mv); | |
270 | 316 |
271 if (frame_mv[NEWMV][ref_frame].as_int == INVALID_MV) | 317 if (frame_mv[NEWMV][ref_frame].as_int == INVALID_MV) |
272 continue; | 318 continue; |
273 | 319 |
| 320 rate_mode = x->inter_mode_cost[mbmi->mode_context[ref_frame]] |
| 321 [INTER_OFFSET(this_mode)]; |
| 322 if (RDCOST(x->rdmult, x->rddiv, rate_mv + rate_mode, 0) > best_rd) |
| 323 continue; |
| 324 |
274 sub_pixel_motion_search(cpi, x, tile, bsize, mi_row, mi_col, | 325 sub_pixel_motion_search(cpi, x, tile, bsize, mi_row, mi_col, |
275 &frame_mv[NEWMV][ref_frame].as_mv); | 326 &frame_mv[NEWMV][ref_frame].as_mv); |
276 } | 327 } |
277 | 328 |
278 if (frame_mv[this_mode][ref_frame].as_int == 0) { | 329 if (this_mode != NEARESTMV) |
279 dist = x->mode_sad[ref_frame][INTER_OFFSET(ZEROMV)]; | 330 if (frame_mv[this_mode][ref_frame].as_int == |
280 } else if (this_mode != NEARESTMV && | 331 frame_mv[NEARESTMV][ref_frame].as_int) |
281 frame_mv[NEARESTMV][ref_frame].as_int == | 332 continue; |
282 frame_mv[this_mode][ref_frame].as_int) { | 333 |
283 dist = x->mode_sad[ref_frame][INTER_OFFSET(NEARESTMV)]; | 334 mbmi->mode = this_mode; |
| 335 mbmi->mv[0].as_int = frame_mv[this_mode][ref_frame].as_int; |
| 336 |
| 337 // Search for the best prediction filter type, when the resulting |
| 338 // motion vector is at sub-pixel accuracy level for luma component, i.e., |
| 339 // the last three bits are all zeros. |
| 340 if ((this_mode == NEWMV || filter_ref == SWITCHABLE) && |
| 341 ((mbmi->mv[0].as_mv.row & 0x07) != 0 || |
| 342 (mbmi->mv[0].as_mv.col & 0x07) != 0)) { |
| 343 int64_t tmp_rdcost1 = INT64_MAX; |
| 344 int64_t tmp_rdcost2 = INT64_MAX; |
| 345 int64_t tmp_rdcost3 = INT64_MAX; |
| 346 int pf_rate[3]; |
| 347 int64_t pf_dist[3]; |
| 348 |
| 349 mbmi->interp_filter = EIGHTTAP; |
| 350 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize); |
| 351 model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[EIGHTTAP], |
| 352 &pf_dist[EIGHTTAP]); |
| 353 tmp_rdcost1 = RDCOST(x->rdmult, x->rddiv, |
| 354 vp9_get_switchable_rate(x) + pf_rate[EIGHTTAP], |
| 355 pf_dist[EIGHTTAP]); |
| 356 |
| 357 mbmi->interp_filter = EIGHTTAP_SHARP; |
| 358 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize); |
| 359 model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[EIGHTTAP_SHARP], |
| 360 &pf_dist[EIGHTTAP_SHARP]); |
| 361 tmp_rdcost2 = RDCOST(x->rdmult, x->rddiv, |
| 362 vp9_get_switchable_rate(x) + pf_rate[EIGHTTAP_SHARP], |
| 363 pf_dist[EIGHTTAP_SHARP]); |
| 364 |
| 365 mbmi->interp_filter = EIGHTTAP_SMOOTH; |
| 366 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize); |
| 367 model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[EIGHTTAP_SMOOTH], |
| 368 &pf_dist[EIGHTTAP_SMOOTH]); |
| 369 tmp_rdcost3 = RDCOST(x->rdmult, x->rddiv, |
| 370 vp9_get_switchable_rate(x) + pf_rate[EIGHTTAP_SMOOTH], |
| 371 pf_dist[EIGHTTAP_SMOOTH]); |
| 372 |
| 373 if (tmp_rdcost2 < tmp_rdcost1) { |
| 374 if (tmp_rdcost2 < tmp_rdcost3) |
| 375 mbmi->interp_filter = EIGHTTAP_SHARP; |
| 376 else |
| 377 mbmi->interp_filter = EIGHTTAP_SMOOTH; |
| 378 } else { |
| 379 if (tmp_rdcost1 < tmp_rdcost3) |
| 380 mbmi->interp_filter = EIGHTTAP; |
| 381 else |
| 382 mbmi->interp_filter = EIGHTTAP_SMOOTH; |
| 383 } |
| 384 |
| 385 rate = pf_rate[mbmi->interp_filter]; |
| 386 dist = pf_dist[mbmi->interp_filter]; |
284 } else { | 387 } else { |
285 mbmi->mode = this_mode; | 388 mbmi->interp_filter = (filter_ref == SWITCHABLE) ? EIGHTTAP: filter_ref; |
286 mbmi->mv[0].as_int = frame_mv[this_mode][ref_frame].as_int; | |
287 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize); | 389 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize); |
288 dist = x->mode_sad[ref_frame][INTER_OFFSET(this_mode)] = | 390 model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist); |
289 cpi->fn_ptr[bsize].sdf(p->src.buf, p->src.stride, | |
290 pd->dst.buf, pd->dst.stride, INT_MAX); | |
291 } | 391 } |
292 | 392 |
293 this_rd = rate + dist; | 393 rate += rate_mv; |
| 394 rate += x->inter_mode_cost[mbmi->mode_context[ref_frame]] |
| 395 [INTER_OFFSET(this_mode)]; |
| 396 this_rd = RDCOST(x->rdmult, x->rddiv, rate, dist); |
294 | 397 |
295 if (this_rd < best_rd) { | 398 if (this_rd < best_rd) { |
296 best_rd = this_rd; | 399 best_rd = this_rd; |
| 400 *returnrate = rate; |
| 401 *returndistortion = dist; |
297 best_mode = this_mode; | 402 best_mode = this_mode; |
| 403 best_pred_filter = mbmi->interp_filter; |
298 best_ref_frame = ref_frame; | 404 best_ref_frame = ref_frame; |
299 } | 405 } |
300 } | 406 } |
301 } | 407 } |
302 | 408 |
303 mbmi->mode = best_mode; | 409 mbmi->mode = best_mode; |
| 410 mbmi->interp_filter = best_pred_filter; |
304 mbmi->ref_frame[0] = best_ref_frame; | 411 mbmi->ref_frame[0] = best_ref_frame; |
305 mbmi->mv[0].as_int = frame_mv[best_mode][best_ref_frame].as_int; | 412 mbmi->mv[0].as_int = frame_mv[best_mode][best_ref_frame].as_int; |
306 xd->mi_8x8[0]->bmi[0].as_mv[0].as_int = mbmi->mv[0].as_int; | 413 xd->mi[0]->bmi[0].as_mv[0].as_int = mbmi->mv[0].as_int; |
307 | 414 |
308 // Perform intra prediction search, if the best SAD is above a certain | 415 // Perform intra prediction search, if the best SAD is above a certain |
309 // threshold. | 416 // threshold. |
310 if (best_rd > inter_mode_thresh) { | 417 if (best_rd > inter_mode_thresh) { |
311 for (this_mode = DC_PRED; this_mode <= DC_PRED; ++this_mode) { | 418 for (this_mode = DC_PRED; this_mode <= DC_PRED; ++this_mode) { |
312 vp9_predict_intra_block(xd, 0, b_width_log2(bsize), | 419 vp9_predict_intra_block(xd, 0, b_width_log2(bsize), |
313 mbmi->tx_size, this_mode, | 420 mbmi->tx_size, this_mode, |
314 &p->src.buf[0], p->src.stride, | 421 &p->src.buf[0], p->src.stride, |
315 &pd->dst.buf[0], pd->dst.stride, 0, 0, 0); | 422 &pd->dst.buf[0], pd->dst.stride, 0, 0, 0); |
316 | 423 |
317 this_rd = cpi->fn_ptr[bsize].sdf(p->src.buf, | 424 model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist); |
318 p->src.stride, | 425 rate += x->mbmode_cost[this_mode]; |
319 pd->dst.buf, | 426 rate += intra_cost_penalty; |
320 pd->dst.stride, INT_MAX); | 427 this_rd = RDCOST(x->rdmult, x->rddiv, rate, dist); |
321 | 428 |
322 if (this_rd + intra_mode_cost < best_rd) { | 429 if (this_rd + intra_mode_cost < best_rd) { |
323 best_rd = this_rd; | 430 best_rd = this_rd; |
| 431 *returnrate = rate; |
| 432 *returndistortion = dist; |
324 mbmi->mode = this_mode; | 433 mbmi->mode = this_mode; |
325 mbmi->ref_frame[0] = INTRA_FRAME; | 434 mbmi->ref_frame[0] = INTRA_FRAME; |
326 mbmi->uv_mode = this_mode; | 435 mbmi->uv_mode = this_mode; |
327 mbmi->mv[0].as_int = INVALID_MV; | 436 mbmi->mv[0].as_int = INVALID_MV; |
328 } | 437 } |
329 } | 438 } |
330 } | 439 } |
| 440 |
331 return INT64_MAX; | 441 return INT64_MAX; |
332 } | 442 } |
OLD | NEW |