Line | Branch | Exec | Source |
---|---|---|---|
1 | /* | ||
2 | * VVC inter prediction | ||
3 | * | ||
4 | * Copyright (C) 2022 Nuo Mi | ||
5 | * | ||
6 | * This file is part of FFmpeg. | ||
7 | * | ||
8 | * FFmpeg is free software; you can redistribute it and/or | ||
9 | * modify it under the terms of the GNU Lesser General Public | ||
10 | * License as published by the Free Software Foundation; either | ||
11 | * version 2.1 of the License, or (at your option) any later version. | ||
12 | * | ||
13 | * FFmpeg is distributed in the hope that it will be useful, | ||
14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | ||
16 | * Lesser General Public License for more details. | ||
17 | * | ||
18 | * You should have received a copy of the GNU Lesser General Public | ||
19 | * License along with FFmpeg; if not, write to the Free Software | ||
20 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | ||
21 | */ | ||
22 | #include "libavutil/frame.h" | ||
23 | |||
24 | #include "data.h" | ||
25 | #include "inter.h" | ||
26 | #include "mvs.h" | ||
27 | #include "refs.h" | ||
28 | |||
29 | // +1 is enough, + 32 for asm alignment | ||
30 | #define PROF_TEMP_OFFSET (MAX_PB_SIZE + 32) | ||
31 | static const int bcw_w_lut[] = {4, 5, 3, 10, -2}; | ||
32 | |||
33 | 20046855 | static void subpic_get_rect(VVCRect *r, const VVCFrame *src_frame, const int subpic_idx, const int is_chroma) | |
34 | { | ||
35 | 20046855 | const VVCSPS *sps = src_frame->sps; | |
36 | 20046855 | const VVCPPS *pps = src_frame->pps; | |
37 | 20046855 | const int hs = sps->hshift[is_chroma]; | |
38 | 20046855 | const int vs = sps->vshift[is_chroma]; | |
39 | |||
40 | 20046855 | r->l = pps->subpic_x[subpic_idx] >> hs; | |
41 | 20046855 | r->t = pps->subpic_y[subpic_idx] >> vs; | |
42 | 20046855 | r->r = r->l + (pps->subpic_width[subpic_idx] >> hs); | |
43 | 20046855 | r->b = r->t + (pps->subpic_height[subpic_idx] >> vs); | |
44 | 20046855 | } | |
45 | |||
46 | // clip to subblock and subpicture process in 8.5.6.3.2 Luma sample interpolation filtering process | ||
47 | 20049096 | static void clip_to_subpic(int *x_off, int *y_off, int *pic_width, int *pic_height, const VVCRect *subpic, const VVCRect *sb, const int dmvr_clip) | |
48 | { | ||
49 |
4/4✓ Branch 0 taken 1783896 times.
✓ Branch 1 taken 18265200 times.
✓ Branch 2 taken 128 times.
✓ Branch 3 taken 1783768 times.
|
20049096 | const int l = dmvr_clip ? FFMIN(FFMAX(subpic->l, sb->l), subpic->r - 1) : subpic->l; |
50 |
4/4✓ Branch 0 taken 1783896 times.
✓ Branch 1 taken 18265200 times.
✓ Branch 2 taken 1066 times.
✓ Branch 3 taken 1782830 times.
|
20049096 | const int t = dmvr_clip ? FFMIN(FFMAX(subpic->t, sb->t), subpic->b - 1) : subpic->t; |
51 |
2/2✓ Branch 0 taken 1783896 times.
✓ Branch 1 taken 18265200 times.
|
20049096 | const int r = dmvr_clip ? FFMAX(FFMIN(subpic->r, sb->r), subpic->l + 1) : subpic->r; |
52 |
2/2✓ Branch 0 taken 1783896 times.
✓ Branch 1 taken 18265200 times.
|
20049096 | const int b = dmvr_clip ? FFMAX(FFMIN(subpic->b, sb->b), subpic->t + 1) : subpic->b; |
53 | |||
54 | 20049096 | *x_off -= l; | |
55 | 20049096 | *y_off -= t; | |
56 | 20049096 | *pic_width = r - l; | |
57 | 20049096 | *pic_height = b - t; | |
58 | 20049096 | } | |
59 | |||
60 | 20044614 | static void emulated_edge_no_wrap(const VVCLocalContext *lc, uint8_t *dst, | |
61 | const uint8_t **src, ptrdiff_t *src_stride, | ||
62 | int x_off, int y_off, const int block_w, const int block_h, | ||
63 | const int extra_before, const int extra_after, | ||
64 | const VVCRect *subpic, const VVCRect *sb, const int dmvr_clip) | ||
65 | { | ||
66 | 20044614 | const VVCFrameContext *fc = lc->fc; | |
67 | 20044614 | const int extra = extra_before + extra_after; | |
68 | int pic_width, pic_height; | ||
69 | |||
70 | 20044614 | *src += y_off * *src_stride + (x_off * (1 << fc->ps.sps->pixel_shift)); | |
71 | |||
72 | 20044614 | clip_to_subpic(&x_off, &y_off, &pic_width, &pic_height, subpic, sb, dmvr_clip); | |
73 | |||
74 |
6/6✓ Branch 0 taken 18260828 times.
✓ Branch 1 taken 1783786 times.
✓ Branch 2 taken 18147789 times.
✓ Branch 3 taken 113039 times.
✓ Branch 4 taken 17920930 times.
✓ Branch 5 taken 226859 times.
|
20044614 | if (dmvr_clip || x_off < extra_before || y_off < extra_before || |
75 |
2/2✓ Branch 0 taken 17790903 times.
✓ Branch 1 taken 130027 times.
|
17920930 | x_off >= pic_width - block_w - extra_after || |
76 |
2/2✓ Branch 0 taken 215596 times.
✓ Branch 1 taken 17575307 times.
|
17790903 | y_off >= pic_height - block_h - extra_after) { |
77 | 2469307 | const int ps = fc->ps.sps->pixel_shift; | |
78 | 2469307 | const ptrdiff_t edge_emu_stride = EDGE_EMU_BUFFER_STRIDE << ps; | |
79 | 2469307 | const int offset = extra_before * *src_stride + (extra_before << ps); | |
80 | 2469307 | const int buf_offset = extra_before * edge_emu_stride + (extra_before << ps); | |
81 | |||
82 | 2469307 | fc->vdsp.emulated_edge_mc(dst, *src - offset, edge_emu_stride, *src_stride, | |
83 | block_w + extra, block_h + extra, x_off - extra_before, y_off - extra_before, | ||
84 | pic_width, pic_height); | ||
85 | |||
86 | 2469307 | *src = dst + buf_offset; | |
87 | 2469307 | *src_stride = edge_emu_stride; | |
88 | } | ||
89 | 20044614 | } | |
90 | |||
91 | 4482 | static void emulated_half(const VVCLocalContext *lc, uint8_t *dst, const ptrdiff_t dst_stride, | |
92 | const uint8_t *src, const ptrdiff_t src_stride, const int ps, | ||
93 | int x_off, int y_off, const int block_w, const int block_h, | ||
94 | const VVCRect *subpic,const VVCRect *half_sb, const int dmvr_clip) | ||
95 | { | ||
96 | 4482 | const VVCFrameContext *fc = lc->fc; | |
97 | int pic_width, pic_height; | ||
98 | |||
99 | 4482 | src += y_off * src_stride + x_off * (1 << ps); | |
100 | |||
101 | 4482 | clip_to_subpic(&x_off, &y_off, &pic_width, &pic_height, subpic, half_sb, dmvr_clip); | |
102 | |||
103 | 4482 | fc->vdsp.emulated_edge_mc(dst, src, dst_stride, src_stride, | |
104 | block_w, block_h, x_off, y_off, pic_width, pic_height); | ||
105 | 4482 | } | |
106 | |||
107 | 9897 | static void sb_set_lr(VVCRect *sb, const int l, const int r) | |
108 | { | ||
109 | 9897 | sb->l = l; | |
110 | 9897 | sb->r = r; | |
111 | 9897 | } | |
112 | |||
113 | 7656 | static void sb_wrap(VVCRect *sb, const int wrap) | |
114 | { | ||
115 | 7656 | sb_set_lr(sb, sb->l + wrap, sb->r + wrap); | |
116 | 7656 | } | |
117 | |||
118 | 20046855 | static void emulated_edge(const VVCLocalContext *lc, uint8_t *dst, | |
119 | const uint8_t **src, ptrdiff_t *src_stride, const VVCFrame *src_frame, | ||
120 | int x_sb, int y_sb, int x_off, int y_off, int block_w, int block_h, const int wrap_enabled, | ||
121 | const int is_chroma, const int extra_before, const int extra_after) | ||
122 | { | ||
123 | 20046855 | const VVCSPS *sps = src_frame->sps; | |
124 | 20046855 | const VVCPPS *pps = src_frame->pps; | |
125 | 20046855 | const int ps = sps->pixel_shift; | |
126 | 20046855 | const int subpic_idx = lc->sc->sh.r->curr_subpic_idx; | |
127 | 20046855 | const int extra = extra_before + extra_after; | |
128 |
4/4✓ Branch 0 taken 18678690 times.
✓ Branch 1 taken 1368165 times.
✓ Branch 2 taken 415676 times.
✓ Branch 3 taken 18263014 times.
|
20046855 | const int dmvr_clip = x_sb != x_off || y_sb != y_off; |
129 | 20046855 | const int dmvr_left = FFMAX(x_off, x_sb) - extra_before; | |
130 | 20046855 | const int dmvr_right = FFMIN(x_off, x_sb) + block_w + extra_after; | |
131 | 20046855 | const int left = x_off - extra_before; | |
132 | 20046855 | const int top = y_off - extra_before; | |
133 | 20046855 | const int pic_width = pps->width >> sps->hshift[is_chroma]; | |
134 | 20046855 | const int wrap = pps->ref_wraparound_offset << (sps->min_cb_log2_size_y - sps->hshift[is_chroma]); | |
135 | 20046855 | const ptrdiff_t dst_stride = EDGE_EMU_BUFFER_STRIDE << ps; | |
136 | 20046855 | VVCRect sb = { x_sb - extra_before, y_sb - extra_before, x_sb + block_w + extra_after, y_sb + block_h + extra_after }; | |
137 | VVCRect subpic; | ||
138 | |||
139 | 20046855 | subpic_get_rect(&subpic, src_frame, subpic_idx, is_chroma); | |
140 | |||
141 |
6/6✓ Branch 0 taken 257536 times.
✓ Branch 1 taken 19789319 times.
✓ Branch 2 taken 251971 times.
✓ Branch 3 taken 5565 times.
✓ Branch 4 taken 247639 times.
✓ Branch 5 taken 4332 times.
|
20046855 | if (!wrap_enabled || (dmvr_left >= 0 && dmvr_right <= pic_width)) { |
142 | 20036958 | emulated_edge_no_wrap(lc, dst, src, src_stride, | |
143 | x_off, y_off, block_w, block_h, extra_before, extra_after, &subpic, &sb, dmvr_clip); | ||
144 | 20044614 | return; | |
145 | } | ||
146 |
2/2✓ Branch 0 taken 4161 times.
✓ Branch 1 taken 5736 times.
|
9897 | if (dmvr_right <= 0) { |
147 | 4161 | sb_wrap(&sb, wrap); | |
148 | 4161 | emulated_edge_no_wrap(lc, dst, src, src_stride, | |
149 | x_off + wrap, y_off, block_w, block_h, extra_before, extra_after, &subpic, &sb, dmvr_clip); | ||
150 | 4161 | return; | |
151 | } | ||
152 |
2/2✓ Branch 0 taken 3495 times.
✓ Branch 1 taken 2241 times.
|
5736 | if (dmvr_left >= pic_width) { |
153 | 3495 | sb_wrap(&sb, -wrap); | |
154 | 3495 | emulated_edge_no_wrap(lc, dst, src, src_stride, | |
155 | x_off - wrap, y_off, block_w, block_h, extra_before, extra_after, &subpic, &sb, dmvr_clip); | ||
156 | 3495 | return; | |
157 | } | ||
158 | |||
159 | 2241 | block_w += extra; | |
160 | 2241 | block_h += extra; | |
161 | |||
162 | // half block are wrapped | ||
163 |
2/2✓ Branch 0 taken 1404 times.
✓ Branch 1 taken 837 times.
|
2241 | if (dmvr_left < 0 ) { |
164 | 1404 | const int w = -left; | |
165 | 1404 | VVCRect half_sb = { sb.l + wrap, sb.t, 0 + wrap, sb.b }; | |
166 | 1404 | emulated_half(lc, dst, dst_stride, *src, *src_stride, ps, | |
167 | left + wrap, top, w, block_h, &subpic, &half_sb, dmvr_clip); | ||
168 | |||
169 | 1404 | sb_set_lr(&half_sb, 0, sb.r); | |
170 | 1404 | emulated_half(lc, dst + (w << ps), dst_stride, *src, *src_stride, ps, | |
171 | 0, top, block_w - w, block_h, &subpic, &half_sb, dmvr_clip); | ||
172 | } else { | ||
173 | 837 | const int w = pic_width - left; | |
174 | 837 | VVCRect half_sb = { sb.l, sb.t, pic_width, sb.b }; | |
175 | 837 | emulated_half(lc, dst, dst_stride, *src, *src_stride, ps, | |
176 | left, top, w, block_h, &subpic, &half_sb, dmvr_clip); | ||
177 | |||
178 | 837 | sb_set_lr(&half_sb, pic_width - wrap, sb.r - wrap); | |
179 | 837 | emulated_half(lc, dst + (w << ps), dst_stride, *src, *src_stride, ps, | |
180 | pic_width - wrap , top, block_w - w, block_h, &subpic, &half_sb, dmvr_clip); | ||
181 | } | ||
182 | |||
183 | 2241 | *src = dst + extra_before * dst_stride + (extra_before << ps); | |
184 | 2241 | *src_stride = dst_stride; | |
185 | } | ||
186 | |||
187 | #define MC_EMULATED_EDGE(dst, src, src_stride, x_off, y_off) \ | ||
188 | emulated_edge(lc, dst, src, src_stride, ref, x_off, y_off, x_off, y_off, block_w, block_h, wrap_enabled, is_chroma, \ | ||
189 | is_chroma ? CHROMA_EXTRA_BEFORE : LUMA_EXTRA_BEFORE, is_chroma ? CHROMA_EXTRA_AFTER : LUMA_EXTRA_AFTER) | ||
190 | |||
191 | #define MC_EMULATED_EDGE_DMVR(dst, src, src_stride, x_sb, y_sb, x_off, y_off) \ | ||
192 | emulated_edge(lc, dst, src, src_stride, ref, x_sb, y_sb, x_off, y_off, block_w, block_h, wrap_enabled, is_chroma, \ | ||
193 | is_chroma ? CHROMA_EXTRA_BEFORE : LUMA_EXTRA_BEFORE, is_chroma ? CHROMA_EXTRA_AFTER : LUMA_EXTRA_AFTER) | ||
194 | |||
195 | #define MC_EMULATED_EDGE_BILINEAR(dst, src, src_stride, x_off, y_off) \ | ||
196 | emulated_edge(lc, dst, src, src_stride, ref, x_off, y_off, x_off, y_off, pred_w, pred_h, wrap_enabled, 0, \ | ||
197 | BILINEAR_EXTRA_BEFORE, BILINEAR_EXTRA_AFTER) | ||
198 | |||
199 | // part of 8.5.6.6 Weighted sample prediction process | ||
200 | 5656789 | static int derive_weight_uni(int *denom, int *wx, int *ox, | |
201 | const VVCLocalContext *lc, const MvField *mvf, const int c_idx) | ||
202 | { | ||
203 | 5656789 | const VVCFrameContext *fc = lc->fc; | |
204 | 5656789 | const VVCPPS *pps = fc->ps.pps; | |
205 | 5656789 | const VVCSH *sh = &lc->sc->sh; | |
206 |
3/4✓ Branch 0 taken 1593390 times.
✓ Branch 1 taken 4063399 times.
✓ Branch 2 taken 1593390 times.
✗ Branch 3 not taken.
|
11313578 | const int weight_flag = (IS_P(sh->r) && pps->r->pps_weighted_pred_flag) || |
207 |
4/4✓ Branch 0 taken 4063399 times.
✓ Branch 1 taken 1593390 times.
✓ Branch 2 taken 22668 times.
✓ Branch 3 taken 4040731 times.
|
5656789 | (IS_B(sh->r) && pps->r->pps_weighted_bipred_flag); |
208 |
2/2✓ Branch 0 taken 22668 times.
✓ Branch 1 taken 5634121 times.
|
5656789 | if (weight_flag) { |
209 | 22668 | const int lx = mvf->pred_flag - PF_L0; | |
210 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 22668 times.
|
22668 | const PredWeightTable *w = pps->r->pps_wp_info_in_ph_flag ? &fc->ps.ph.pwt : &sh->pwt; |
211 | |||
212 | 22668 | *denom = w->log2_denom[c_idx > 0]; | |
213 | 22668 | *wx = w->weight[lx][c_idx][mvf->ref_idx[lx]]; | |
214 | 22668 | *ox = w->offset[lx][c_idx][mvf->ref_idx[lx]]; | |
215 | } | ||
216 | 5656789 | return weight_flag; | |
217 | } | ||
218 | |||
219 | // part of 8.5.6.6 Weighted sample prediction process | ||
220 | 6139018 | static int derive_weight(int *denom, int *w0, int *w1, int *o0, int *o1, | |
221 | const VVCLocalContext *lc, const MvField *mvf, const int c_idx, const int dmvr_flag) | ||
222 | { | ||
223 | 6139018 | const VVCFrameContext *fc = lc->fc; | |
224 | 6139018 | const VVCPPS *pps = fc->ps.pps; | |
225 | 6139018 | const VVCSH *sh = &lc->sc->sh; | |
226 | 6139018 | const int bcw_idx = mvf->bcw_idx; | |
227 |
1/4✗ Branch 0 not taken.
✓ Branch 1 taken 6139018 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
12278036 | const int weight_flag = (IS_P(sh->r) && pps->r->pps_weighted_pred_flag) || |
228 |
4/6✓ Branch 0 taken 6139018 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 16347 times.
✓ Branch 3 taken 6122671 times.
✓ Branch 4 taken 16347 times.
✗ Branch 5 not taken.
|
6139018 | (IS_B(sh->r) && pps->r->pps_weighted_bipred_flag && !dmvr_flag); |
229 |
8/8✓ Branch 0 taken 6122671 times.
✓ Branch 1 taken 16347 times.
✓ Branch 2 taken 500326 times.
✓ Branch 3 taken 5622345 times.
✓ Branch 4 taken 500326 times.
✓ Branch 5 taken 16347 times.
✓ Branch 6 taken 3703 times.
✓ Branch 7 taken 496623 times.
|
6139018 | if ((!weight_flag && !bcw_idx) || (bcw_idx && lc->cu->ciip_flag)) |
230 | 5626048 | return 0; | |
231 | |||
232 |
2/2✓ Branch 0 taken 496623 times.
✓ Branch 1 taken 16347 times.
|
512970 | if (bcw_idx) { |
233 | 496623 | *denom = 2; | |
234 | 496623 | *w1 = bcw_w_lut[bcw_idx]; | |
235 | 496623 | *w0 = 8 - *w1; | |
236 | 496623 | *o0 = *o1 = 0; | |
237 | } else { | ||
238 | 16347 | const VVCPPS *pps = fc->ps.pps; | |
239 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 16347 times.
|
16347 | const PredWeightTable *w = pps->r->pps_wp_info_in_ph_flag ? &fc->ps.ph.pwt : &sh->pwt; |
240 | |||
241 | 16347 | *denom = w->log2_denom[c_idx > 0]; | |
242 | 16347 | *w0 = w->weight[L0][c_idx][mvf->ref_idx[L0]]; | |
243 | 16347 | *w1 = w->weight[L1][c_idx][mvf->ref_idx[L1]]; | |
244 | 16347 | *o0 = w->offset[L0][c_idx][mvf->ref_idx[L0]]; | |
245 | 16347 | *o1 = w->offset[L1][c_idx][mvf->ref_idx[L1]]; | |
246 | } | ||
247 | 512970 | return 1; | |
248 | } | ||
249 | |||
250 | #define INTER_FILTER(t, frac) (is_chroma ? ff_vvc_inter_chroma_filters[t][frac] : ff_vvc_inter_luma_filters[t][frac]) | ||
251 | |||
252 | 138596 | static void mc(VVCLocalContext *lc, int16_t *dst, const VVCFrame *ref, const Mv *mv, | |
253 | int x_off, int y_off, const int block_w, const int block_h, const int c_idx) | ||
254 | { | ||
255 | 138596 | const VVCFrameContext *fc = lc->fc; | |
256 | 138596 | const PredictionUnit *pu = &lc->cu->pu; | |
257 | 138596 | const uint8_t *src = ref->frame->data[c_idx]; | |
258 | 138596 | ptrdiff_t src_stride = ref->frame->linesize[c_idx]; | |
259 | 138596 | const int is_chroma = !!c_idx; | |
260 | 138596 | const int hs = fc->ps.sps->hshift[c_idx]; | |
261 | 138596 | const int vs = fc->ps.sps->vshift[c_idx]; | |
262 | 138596 | const int idx = av_log2(block_w) - 1; | |
263 | 138596 | const intptr_t mx = av_zero_extend(mv->x, 4 + hs) << (is_chroma - hs); | |
264 | 138596 | const intptr_t my = av_zero_extend(mv->y, 4 + vs) << (is_chroma - vs); | |
265 |
4/4✓ Branch 0 taken 49240 times.
✓ Branch 1 taken 89356 times.
✓ Branch 2 taken 39 times.
✓ Branch 3 taken 49201 times.
|
138596 | const int hpel_if_idx = (is_chroma || pu->merge_gpm_flag) ? 0 : pu->mi.hpel_if_idx; |
266 |
2/2✓ Branch 0 taken 89356 times.
✓ Branch 1 taken 49240 times.
|
138596 | const int8_t *hf = INTER_FILTER(hpel_if_idx, mx); |
267 |
2/2✓ Branch 0 taken 89356 times.
✓ Branch 1 taken 49240 times.
|
138596 | const int8_t *vf = INTER_FILTER(hpel_if_idx, my); |
268 | 138596 | const int wrap_enabled = fc->ps.pps->r->pps_ref_wraparound_enabled_flag; | |
269 | |||
270 | 138596 | x_off += mv->x >> (4 + hs); | |
271 | 138596 | y_off += mv->y >> (4 + vs); | |
272 | |||
273 |
4/4✓ Branch 0 taken 89356 times.
✓ Branch 1 taken 49240 times.
✓ Branch 2 taken 89356 times.
✓ Branch 3 taken 49240 times.
|
138596 | MC_EMULATED_EDGE(lc->edge_emu_buffer, &src, &src_stride, x_off, y_off); |
274 | 138596 | fc->vvcdsp.inter.put[is_chroma][idx][!!my][!!mx](dst, src, src_stride, block_h, hf, vf, block_w); | |
275 | 138596 | } | |
276 | |||
277 | 3345036 | static void mc_uni(VVCLocalContext *lc, uint8_t *dst, const ptrdiff_t dst_stride, | |
278 | const VVCFrame *ref, const MvField *mvf, int x_off, int y_off, const int block_w, const int block_h, | ||
279 | const int c_idx) | ||
280 | { | ||
281 | 3345036 | const VVCFrameContext *fc = lc->fc; | |
282 | 3345036 | const PredictionUnit *pu = &lc->cu->pu; | |
283 | 3345036 | const uint8_t *src = ref->frame->data[c_idx]; | |
284 | 3345036 | ptrdiff_t src_stride = ref->frame->linesize[c_idx]; | |
285 | 3345036 | const int lx = mvf->pred_flag - PF_L0; | |
286 | 3345036 | const int hs = fc->ps.sps->hshift[c_idx]; | |
287 | 3345036 | const int vs = fc->ps.sps->vshift[c_idx]; | |
288 | 3345036 | const int idx = av_log2(block_w) - 1; | |
289 | 3345036 | const Mv *mv = &mvf->mv[lx]; | |
290 | 3345036 | const int is_chroma = !!c_idx; | |
291 | 3345036 | const intptr_t mx = av_zero_extend(mv->x, 4 + hs) << (is_chroma - hs); | |
292 | 3345036 | const intptr_t my = av_zero_extend(mv->y, 4 + vs) << (is_chroma - vs); | |
293 |
2/2✓ Branch 0 taken 617678 times.
✓ Branch 1 taken 2727358 times.
|
3345036 | const int hpel_if_idx = is_chroma ? 0 : pu->mi.hpel_if_idx; |
294 |
2/2✓ Branch 0 taken 2727358 times.
✓ Branch 1 taken 617678 times.
|
3345036 | const int8_t *hf = INTER_FILTER(hpel_if_idx, mx); |
295 |
2/2✓ Branch 0 taken 2727358 times.
✓ Branch 1 taken 617678 times.
|
3345036 | const int8_t *vf = INTER_FILTER(hpel_if_idx, my); |
296 | 3345036 | const int wrap_enabled = fc->ps.pps->r->pps_ref_wraparound_enabled_flag; | |
297 | int denom, wx, ox; | ||
298 | |||
299 | 3345036 | x_off += mv->x >> (4 + hs); | |
300 | 3345036 | y_off += mv->y >> (4 + vs); | |
301 | |||
302 |
4/4✓ Branch 0 taken 2727358 times.
✓ Branch 1 taken 617678 times.
✓ Branch 2 taken 2727358 times.
✓ Branch 3 taken 617678 times.
|
3345036 | MC_EMULATED_EDGE(lc->edge_emu_buffer, &src, &src_stride, x_off, y_off); |
303 |
2/2✓ Branch 1 taken 17244 times.
✓ Branch 2 taken 3327792 times.
|
3345036 | if (derive_weight_uni(&denom, &wx, &ox, lc, mvf, c_idx)) { |
304 | 17244 | fc->vvcdsp.inter.put_uni_w[is_chroma][idx][!!my][!!mx](dst, dst_stride, src, src_stride, | |
305 | block_h, denom, wx, ox, hf, vf, block_w); | ||
306 | } else { | ||
307 | 3327792 | fc->vvcdsp.inter.put_uni[is_chroma][idx][!!my][!!mx](dst, dst_stride, src, src_stride, | |
308 | block_h, hf, vf, block_w); | ||
309 | } | ||
310 | 3345036 | } | |
311 | |||
312 | 5269696 | static void mc_bi(VVCLocalContext *lc, uint8_t *dst, const ptrdiff_t dst_stride, | |
313 | const VVCFrame *ref0, const VVCFrame *ref1, const MvField *mvf, const MvField *orig_mv, | ||
314 | const int x_off, const int y_off, const int block_w, const int block_h, const int c_idx, | ||
315 | const int sb_bdof_flag) | ||
316 | { | ||
317 | 5269696 | const VVCFrameContext *fc = lc->fc; | |
318 | 5269696 | const PredictionUnit *pu = &lc->cu->pu; | |
319 | 5269696 | const int hs = fc->ps.sps->hshift[c_idx]; | |
320 | 5269696 | const int vs = fc->ps.sps->vshift[c_idx]; | |
321 | 5269696 | const int idx = av_log2(block_w) - 1; | |
322 | 5269696 | const VVCFrame *refs[] = { ref0, ref1 }; | |
323 | 5269696 | int16_t *tmp[] = { lc->tmp + sb_bdof_flag * PROF_TEMP_OFFSET, lc->tmp1 + sb_bdof_flag * PROF_TEMP_OFFSET }; | |
324 | int denom, w0, w1, o0, o1; | ||
325 | 5269696 | const int weight_flag = derive_weight(&denom, &w0, &w1, &o0, &o1, lc, mvf, c_idx, pu->dmvr_flag); | |
326 | 5269696 | const int is_chroma = !!c_idx; | |
327 |
2/2✓ Branch 0 taken 1517616 times.
✓ Branch 1 taken 3752080 times.
|
5269696 | const int hpel_if_idx = is_chroma ? 0 : pu->mi.hpel_if_idx; |
328 | |||
329 |
2/2✓ Branch 0 taken 10539392 times.
✓ Branch 1 taken 5269696 times.
|
15809088 | for (int i = L0; i <= L1; i++) { |
330 | 10539392 | const Mv *mv = mvf->mv + i; | |
331 | 10539392 | const int mx = av_zero_extend(mv->x, 4 + hs) << (is_chroma - hs); | |
332 | 10539392 | const int my = av_zero_extend(mv->y, 4 + vs) << (is_chroma - vs); | |
333 | 10539392 | const int ox = x_off + (mv->x >> (4 + hs)); | |
334 | 10539392 | const int oy = y_off + (mv->y >> (4 + vs)); | |
335 | 10539392 | const VVCFrame *ref = refs[i]; | |
336 | 10539392 | ptrdiff_t src_stride = ref->frame->linesize[c_idx]; | |
337 | 10539392 | const uint8_t *src = ref->frame->data[c_idx]; | |
338 |
2/2✓ Branch 0 taken 7504160 times.
✓ Branch 1 taken 3035232 times.
|
10539392 | const int8_t *hf = INTER_FILTER(hpel_if_idx, mx); |
339 |
2/2✓ Branch 0 taken 7504160 times.
✓ Branch 1 taken 3035232 times.
|
10539392 | const int8_t *vf = INTER_FILTER(hpel_if_idx, my); |
340 | 10539392 | const int wrap_enabled = fc->ps.pps->r->pps_ref_wraparound_enabled_flag; | |
341 | |||
342 |
2/2✓ Branch 0 taken 5885324 times.
✓ Branch 1 taken 4654068 times.
|
10539392 | if (pu->dmvr_flag) { |
343 | 5885324 | const int x_sb = x_off + (orig_mv->mv[i].x >> (4 + hs)); | |
344 | 5885324 | const int y_sb = y_off + (orig_mv->mv[i].y >> (4 + vs)); | |
345 | |||
346 |
4/4✓ Branch 0 taken 3912592 times.
✓ Branch 1 taken 1972732 times.
✓ Branch 2 taken 3912592 times.
✓ Branch 3 taken 1972732 times.
|
5885324 | MC_EMULATED_EDGE_DMVR(lc->edge_emu_buffer, &src, &src_stride, x_sb, y_sb, ox, oy); |
347 | } else { | ||
348 |
4/4✓ Branch 0 taken 3591568 times.
✓ Branch 1 taken 1062500 times.
✓ Branch 2 taken 3591568 times.
✓ Branch 3 taken 1062500 times.
|
4654068 | MC_EMULATED_EDGE(lc->edge_emu_buffer, &src, &src_stride, ox, oy); |
349 | } | ||
350 | 10539392 | fc->vvcdsp.inter.put[is_chroma][idx][!!my][!!mx](tmp[i], src, src_stride, block_h, hf, vf, block_w); | |
351 |
2/2✓ Branch 0 taken 700286 times.
✓ Branch 1 taken 9839106 times.
|
10539392 | if (sb_bdof_flag) |
352 | 700286 | fc->vvcdsp.inter.bdof_fetch_samples(tmp[i], src, src_stride, mx, my, block_w, block_h); | |
353 | } | ||
354 |
2/2✓ Branch 0 taken 350143 times.
✓ Branch 1 taken 4919553 times.
|
5269696 | if (sb_bdof_flag) |
355 | 350143 | fc->vvcdsp.inter.apply_bdof(dst, dst_stride, tmp[L0], tmp[L1], block_w, block_h); | |
356 |
2/2✓ Branch 0 taken 309902 times.
✓ Branch 1 taken 4609651 times.
|
4919553 | else if (weight_flag) |
357 | 309902 | fc->vvcdsp.inter.w_avg(dst, dst_stride, tmp[L0], tmp[L1], block_w, block_h, denom, w0, w1, o0, o1); | |
358 | else | ||
359 | 4609651 | fc->vvcdsp.inter.avg(dst, dst_stride, tmp[L0], tmp[L1], block_w, block_h); | |
360 | 5269696 | } | |
361 | |||
362 | 65406 | static const int8_t* inter_filter_scaled(const int scale, const int is_chroma, const int is_affine) | |
363 | { | ||
364 | #define SCALE_THRESHOLD_1 20480 | ||
365 | #define SCALE_THRESHOLD_2 28672 | ||
366 | |||
367 | 65406 | const int i = (scale > SCALE_THRESHOLD_2) + (scale > SCALE_THRESHOLD_1); | |
368 | |||
369 |
2/2✓ Branch 0 taken 40882 times.
✓ Branch 1 taken 24524 times.
|
65406 | if (!is_chroma) { |
370 |
2/2✓ Branch 0 taken 2722 times.
✓ Branch 1 taken 38160 times.
|
40882 | if (!is_affine) |
371 | 2722 | return &ff_vvc_inter_luma_filters[i + !!i][0][0]; //hpel 1 is not needed for scaled | |
372 | 38160 | return &ff_vvc_inter_luma_filters[VVC_INTER_LUMA_FILTER_TYPE_AFFINE + i][0][0]; | |
373 | } | ||
374 | |||
375 | 24524 | return &ff_vvc_inter_chroma_filters[i][0][0]; | |
376 | } | ||
377 | #define INTER_FILTER_SCALED(scale) inter_filter_scaled(scale, is_chroma, is_affine) | ||
378 | |||
379 | #define SCALED_CHROMA_ADDIN(scale, collocated_flag) (is_chroma ? (collocated_flag ? 0 : 8 * (scale - (1 << 14))) : 0) | ||
380 | #define SCALED_REF_SB(off, scaling_off, ref_mv, scale, add, shift) ((((off - (scaling_off << shift)) << (4 + shift)) + ref_mv) * scale + add) | ||
381 | #define SCALED_REF(ref_sb, offset, shift) (FFSIGN(ref_sb) * ((FFABS(ref_sb) + (128 << is_chroma)) >> (8 + is_chroma)) + (offset << (10 - shift)) + (32 >> is_chroma)) | ||
382 | #define SCALED_STEP(scale) ((scale + 8) >> 4) | ||
383 | |||
384 | 32703 | static void scaled_ref_pos_and_step(const VVCLocalContext *lc, const VVCRefPic *refp, const Mv *mv, const int x_off, const int y_off, const int c_idx, | |
385 | int *x, int *y, int *dx, int *dy) | ||
386 | { | ||
387 | 32703 | const VVCFrameContext *fc = lc->fc; | |
388 | 32703 | const VVCSPS *sps = fc->ps.sps; | |
389 | 32703 | const int is_chroma = !!c_idx; | |
390 | 32703 | const int hs = sps->hshift[c_idx]; | |
391 | 32703 | const int vs = sps->vshift[c_idx]; | |
392 | 32703 | const int left_offset = fc->ref->scaling_win.left_offset; | |
393 | 32703 | const int top_offset = fc->ref->scaling_win.top_offset; | |
394 |
3/4✓ Branch 0 taken 12262 times.
✓ Branch 1 taken 20441 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 12262 times.
|
32703 | const int addx = SCALED_CHROMA_ADDIN(refp->scale[0], sps->r->sps_chroma_horizontal_collocated_flag); |
395 |
3/4✓ Branch 0 taken 12262 times.
✓ Branch 1 taken 20441 times.
✓ Branch 2 taken 12262 times.
✗ Branch 3 not taken.
|
32703 | const int addy = SCALED_CHROMA_ADDIN(refp->scale[1], sps->r->sps_chroma_vertical_collocated_flag); |
396 | 32703 | const int refx_sb = SCALED_REF_SB(x_off, left_offset, mv->x, refp->scale[0], addx, hs); | |
397 | 32703 | const int refy_sb = SCALED_REF_SB(y_off, top_offset, mv->y, refp->scale[1], addy, vs); | |
398 | |||
399 |
2/2✓ Branch 0 taken 32637 times.
✓ Branch 1 taken 66 times.
|
32703 | *x = SCALED_REF(refx_sb, left_offset, hs); |
400 |
2/2✓ Branch 0 taken 32535 times.
✓ Branch 1 taken 168 times.
|
32703 | *y = SCALED_REF(refy_sb, top_offset, vs); |
401 | 32703 | *dx = SCALED_STEP(refp->scale[0]); | |
402 | 32703 | *dy = SCALED_STEP(refp->scale[1]); | |
403 | 32703 | } | |
404 | |||
405 | 32703 | static void emulated_edge_scaled(VVCLocalContext *lc, const uint8_t **src, ptrdiff_t *src_stride, int *src_height, | |
406 | const VVCFrame *ref, const int x, const int y, const int dx, const int dy, const int w, const int h, const int is_chroma) | ||
407 | { | ||
408 | 32703 | const int x_off = SCALED_INT(x); | |
409 | 32703 | const int y_off = SCALED_INT(y); | |
410 | 32703 | const int x_end = SCALED_INT(x + w * dx); | |
411 | 32703 | const int y_end = SCALED_INT(y + h * dy); | |
412 | 32703 | const int x_last = SCALED_INT(x + (w - 1) * dx); | |
413 | 32703 | const int y_last = SCALED_INT(y + (h - 1) * dy); | |
414 | 32703 | const int block_w = x_end - x_off + (x_end == x_last); | |
415 | 32703 | const int block_h = *src_height = y_end - y_off + (y_end == y_last); | |
416 | 32703 | const int wrap_enabled = 0; | |
417 | |||
418 |
4/4✓ Branch 0 taken 12262 times.
✓ Branch 1 taken 20441 times.
✓ Branch 2 taken 12262 times.
✓ Branch 3 taken 20441 times.
|
32703 | MC_EMULATED_EDGE(lc->edge_emu_buffer, src, src_stride, x_off, y_off); |
419 | 32703 | } | |
420 | |||
421 | 5946 | static void mc_scaled(VVCLocalContext *lc, int16_t *dst, const VVCRefPic *refp, const Mv *mv, | |
422 | int x_off, int y_off, const int block_w, const int block_h, const int c_idx) | ||
423 | { | ||
424 | 5946 | const VVCFrameContext *fc = lc->fc; | |
425 | 5946 | const PredictionUnit *pu = &lc->cu->pu; | |
426 | 5946 | const uint8_t *src = refp->ref->frame->data[c_idx]; | |
427 | 5946 | ptrdiff_t src_stride = refp->ref->frame->linesize[c_idx]; | |
428 | 5946 | const int is_affine = pu->inter_affine_flag; | |
429 | 5946 | const int is_chroma = !!c_idx; | |
430 | 5946 | const int idx = av_log2(block_w) - 1; | |
431 | 5946 | const int8_t *hf = INTER_FILTER_SCALED(refp->scale[0]); | |
432 | 5946 | const int8_t *vf = INTER_FILTER_SCALED(refp->scale[1]); | |
433 | int x, y, dx, dy, src_height; | ||
434 | |||
435 | 5946 | scaled_ref_pos_and_step(lc, refp, mv, x_off, y_off, c_idx, &x, &y, &dx, &dy); | |
436 | 5946 | emulated_edge_scaled(lc, &src, &src_stride, &src_height, refp->ref, x, y, dx, dy, block_w, block_h, is_chroma); | |
437 | 5946 | fc->vvcdsp.inter.put_scaled[is_chroma][idx](dst, src, src_stride, src_height, x, y, dx, dy, block_h, hf, vf, block_w); | |
438 | 5946 | } | |
439 | |||
440 | 26757 | static void mc_uni_scaled(VVCLocalContext *lc, uint8_t *dst, const ptrdiff_t dst_stride, const VVCRefPic *refp, | |
441 | const MvField *mvf, const int x_off, const int y_off, const int block_w, const int block_h, const int c_idx) | ||
442 | { | ||
443 | 26757 | const VVCFrameContext *fc = lc->fc; | |
444 | 26757 | const PredictionUnit *pu = &lc->cu->pu; | |
445 | 26757 | const uint8_t *src = refp->ref->frame->data[c_idx]; | |
446 | 26757 | ptrdiff_t src_stride = refp->ref->frame->linesize[c_idx]; | |
447 | 26757 | const int lx = mvf->pred_flag - PF_L0; | |
448 | 26757 | const Mv *mv = &mvf->mv[lx]; | |
449 | 26757 | const int is_affine = pu->inter_affine_flag; | |
450 | 26757 | const int is_chroma = !!c_idx; | |
451 | 26757 | const int idx = av_log2(block_w) - 1; | |
452 | 26757 | const int8_t *hf = INTER_FILTER_SCALED(refp->scale[0]); | |
453 | 26757 | const int8_t *vf = INTER_FILTER_SCALED(refp->scale[1]); | |
454 | int denom, wx, ox, x, y, dx, dy, src_height; | ||
455 | |||
456 | 26757 | scaled_ref_pos_and_step(lc, refp, mv, x_off, y_off, c_idx, &x, &y, &dx, &dy); | |
457 | 26757 | emulated_edge_scaled(lc, &src, &src_stride, &src_height, refp->ref, x, y, dx, dy, block_w, block_h, is_chroma); | |
458 | |||
459 |
1/2✗ Branch 1 not taken.
✓ Branch 2 taken 26757 times.
|
26757 | if (derive_weight_uni(&denom, &wx, &ox, lc, mvf, c_idx)) { |
460 | ✗ | fc->vvcdsp.inter.put_uni_w_scaled[is_chroma][idx](dst, dst_stride, src, src_stride, src_height, | |
461 | x, y, dx, dy, block_h, denom, wx, ox, hf, vf, block_w); | ||
462 | } else { | ||
463 | 26757 | fc->vvcdsp.inter.put_uni_scaled[is_chroma][idx](dst, dst_stride, src, src_stride, src_height, | |
464 | x, y, dx, dy, block_h, hf, vf, block_w); | ||
465 | } | ||
466 | 26757 | } | |
467 | |||
468 | 1086 | static void mc_bi_scaled(VVCLocalContext *lc, uint8_t *dst, const ptrdiff_t dst_stride, | |
469 | const VVCRefPic *refp0, const VVCRefPic *refp1, const MvField *mvf, | ||
470 | const int x_off, const int y_off, const int block_w, const int block_h, const int c_idx) | ||
471 | { | ||
472 | int denom, w0, w1, o0, o1; | ||
473 | 1086 | const VVCFrameContext *fc = lc->fc; | |
474 | 1086 | const int weight_flag = derive_weight(&denom, &w0, &w1, &o0, &o1, lc, mvf, c_idx, lc->cu->pu.dmvr_flag); | |
475 | 1086 | const VVCRefPic *refps[] = { refp0, refp1 }; | |
476 | 1086 | int16_t *tmp[] = { lc->tmp, lc->tmp1 }; | |
477 | |||
478 |
2/2✓ Branch 0 taken 2172 times.
✓ Branch 1 taken 1086 times.
|
3258 | for (int i = L0; i <= L1; i++) { |
479 | 2172 | const Mv *mv = mvf->mv + i; | |
480 | 2172 | const VVCRefPic *refp = refps[i]; | |
481 | |||
482 |
2/2✓ Branch 0 taken 2055 times.
✓ Branch 1 taken 117 times.
|
2172 | if (refp->is_scaled) |
483 | 2055 | mc_scaled(lc, tmp[i], refp, mv, x_off, y_off, block_w, block_h, c_idx); | |
484 | else | ||
485 | 117 | mc(lc, tmp[i], refp->ref, mv, x_off, y_off, block_w, block_h, c_idx); | |
486 | } | ||
487 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 1086 times.
|
1086 | if (weight_flag) |
488 | ✗ | fc->vvcdsp.inter.w_avg(dst, dst_stride, tmp[L0], tmp[L1], block_w, block_h, denom, w0, w1, o0, o1); | |
489 | else | ||
490 | 1086 | fc->vvcdsp.inter.avg(dst, dst_stride, tmp[L0], tmp[L1], block_w, block_h); | |
491 | 1086 | } | |
492 | |||
493 | 2284996 | static void luma_prof_uni(VVCLocalContext *lc, uint8_t *dst, const ptrdiff_t dst_stride, | |
494 | const VVCFrame *ref, const MvField *mvf, int x_off, int y_off, const int block_w, const int block_h, | ||
495 | const int cb_prof_flag, const int16_t *diff_mv_x, const int16_t *diff_mv_y) | ||
496 | { | ||
497 | 2284996 | const VVCFrameContext *fc = lc->fc; | |
498 | 2284996 | const uint8_t *src = ref->frame->data[LUMA]; | |
499 | 2284996 | ptrdiff_t src_stride = ref->frame->linesize[LUMA]; | |
500 | 2284996 | uint16_t *prof_tmp = lc->tmp + PROF_TEMP_OFFSET; | |
501 | 2284996 | const int idx = av_log2(block_w) - 1; | |
502 | 2284996 | const int lx = mvf->pred_flag - PF_L0; | |
503 | 2284996 | const Mv *mv = mvf->mv + lx; | |
504 | 2284996 | const int mx = mv->x & 0xf; | |
505 | 2284996 | const int my = mv->y & 0xf; | |
506 | 2284996 | const int8_t *hf = ff_vvc_inter_luma_filters[VVC_INTER_LUMA_FILTER_TYPE_AFFINE][mx]; | |
507 | 2284996 | const int8_t *vf = ff_vvc_inter_luma_filters[VVC_INTER_LUMA_FILTER_TYPE_AFFINE][my]; | |
508 | int denom, wx, ox; | ||
509 | 2284996 | const int weight_flag = derive_weight_uni(&denom, &wx, &ox, lc, mvf, LUMA); | |
510 | 2284996 | const int wrap_enabled = fc->ps.pps->r->pps_ref_wraparound_enabled_flag; | |
511 | 2284996 | const int is_chroma = 0; | |
512 | |||
513 | 2284996 | x_off += mv->x >> 4; | |
514 | 2284996 | y_off += mv->y >> 4; | |
515 | |||
516 |
2/4✗ Branch 0 not taken.
✓ Branch 1 taken 2284996 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 2284996 times.
|
2284996 | MC_EMULATED_EDGE(lc->edge_emu_buffer, &src, &src_stride, x_off, y_off); |
517 |
2/2✓ Branch 0 taken 2128360 times.
✓ Branch 1 taken 156636 times.
|
2284996 | if (cb_prof_flag) { |
518 | 2128360 | fc->vvcdsp.inter.put[LUMA][idx][!!my][!!mx](prof_tmp, src, src_stride, AFFINE_MIN_BLOCK_SIZE, hf, vf, AFFINE_MIN_BLOCK_SIZE); | |
519 | 2128360 | fc->vvcdsp.inter.fetch_samples(prof_tmp, src, src_stride, mx, my); | |
520 |
2/2✓ Branch 0 taken 2123076 times.
✓ Branch 1 taken 5284 times.
|
2128360 | if (!weight_flag) |
521 | 2123076 | fc->vvcdsp.inter.apply_prof_uni(dst, dst_stride, prof_tmp, diff_mv_x, diff_mv_y); | |
522 | else | ||
523 | 5284 | fc->vvcdsp.inter.apply_prof_uni_w(dst, dst_stride, prof_tmp, diff_mv_x, diff_mv_y, denom, wx, ox); | |
524 | } else { | ||
525 |
2/2✓ Branch 0 taken 156496 times.
✓ Branch 1 taken 140 times.
|
156636 | if (!weight_flag) |
526 | 156496 | fc->vvcdsp.inter.put_uni[LUMA][idx][!!my][!!mx](dst, dst_stride, src, src_stride, block_h, hf, vf, block_w); | |
527 | else | ||
528 | 140 | fc->vvcdsp.inter.put_uni_w[LUMA][idx][!!my][!!mx](dst, dst_stride, src, src_stride, block_h, denom, wx, ox, hf, vf, block_w); | |
529 | } | ||
530 | 2284996 | } | |
531 | |||
532 | 1733400 | static void luma_prof(VVCLocalContext *lc, int16_t *dst, const VVCFrame *ref, | |
533 | const Mv *mv , const int x_off, const int y_off, const int block_w, const int block_h, const int lx) | ||
534 | { | ||
535 | 1733400 | const VVCFrameContext *fc = lc->fc; | |
536 | 1733400 | const PredictionUnit *pu = &lc->cu->pu; | |
537 | 1733400 | const int mx = mv->x & 0xf; | |
538 | 1733400 | const int my = mv->y & 0xf; | |
539 | 1733400 | const int ox = x_off + (mv->x >> 4); | |
540 | 1733400 | const int oy = y_off + (mv->y >> 4); | |
541 | 1733400 | const int idx = av_log2(block_w) - 1; | |
542 | 1733400 | const int is_chroma = 0; | |
543 | 1733400 | uint16_t *prof_tmp = lc->tmp2 + PROF_TEMP_OFFSET; | |
544 | 1733400 | ptrdiff_t src_stride = ref->frame->linesize[LUMA]; | |
545 | 1733400 | const uint8_t *src = ref->frame->data[LUMA]; | |
546 | 1733400 | const int8_t *hf = ff_vvc_inter_luma_filters[VVC_INTER_LUMA_FILTER_TYPE_AFFINE][mx]; | |
547 | 1733400 | const int8_t *vf = ff_vvc_inter_luma_filters[VVC_INTER_LUMA_FILTER_TYPE_AFFINE][my]; | |
548 | 1733400 | const int wrap_enabled = fc->ps.pps->r->pps_ref_wraparound_enabled_flag; | |
549 | |||
550 |
2/4✗ Branch 0 not taken.
✓ Branch 1 taken 1733400 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 1733400 times.
|
1733400 | MC_EMULATED_EDGE(lc->edge_emu_buffer, &src, &src_stride, ox, oy); |
551 |
2/2✓ Branch 0 taken 366992 times.
✓ Branch 1 taken 1366408 times.
|
1733400 | if (!pu->cb_prof_flag[lx]) { |
552 | 366992 | fc->vvcdsp.inter.put[LUMA][idx][!!my][!!mx](dst, src, src_stride, block_h, hf, vf, block_w); | |
553 | } else { | ||
554 | 1366408 | fc->vvcdsp.inter.put[LUMA][idx][!!my][!!mx](prof_tmp, src, src_stride, AFFINE_MIN_BLOCK_SIZE, hf, vf, AFFINE_MIN_BLOCK_SIZE); | |
555 | 1366408 | fc->vvcdsp.inter.fetch_samples(prof_tmp, src, src_stride, mx, my); | |
556 | 1366408 | fc->vvcdsp.inter.apply_prof(dst, prof_tmp, pu->diff_mv_x[lx], pu->diff_mv_y[lx]); | |
557 | } | ||
558 | 1733400 | } | |
559 | |||
560 | 868236 | static void luma_prof_bi(VVCLocalContext *lc, uint8_t *dst, const ptrdiff_t dst_stride, | |
561 | const VVCRefPic *ref0, const VVCRefPic *ref1, const MvField *mvf, const int x_off, const int y_off, | ||
562 | const int block_w, const int block_h) | ||
563 | { | ||
564 | 868236 | const VVCFrameContext *fc = lc->fc; | |
565 | 868236 | const VVCRefPic *refps[] = { ref0, ref1 }; | |
566 | 868236 | int16_t *tmp[] = { lc->tmp, lc->tmp1 }; | |
567 | int denom, w0, w1, o0, o1; | ||
568 | 868236 | const int weight_flag = derive_weight(&denom, &w0, &w1, &o0, &o1, lc, mvf, LUMA, 0); | |
569 | |||
570 |
2/2✓ Branch 0 taken 1736472 times.
✓ Branch 1 taken 868236 times.
|
2604708 | for (int i = L0; i <= L1; i++) { |
571 | 1736472 | const VVCRefPic *refp = refps[i]; | |
572 | 1736472 | const Mv *mv = mvf->mv + i; | |
573 | |||
574 |
2/2✓ Branch 0 taken 3072 times.
✓ Branch 1 taken 1733400 times.
|
1736472 | if (refp->is_scaled) |
575 | 3072 | mc_scaled(lc, tmp[i], refp, mv, x_off, y_off, block_w, block_h, LUMA); | |
576 | else | ||
577 | 1733400 | luma_prof(lc, tmp[i], refp->ref, mv, x_off, y_off, block_w, block_h, i); | |
578 | } | ||
579 | |||
580 |
2/2✓ Branch 0 taken 203068 times.
✓ Branch 1 taken 665168 times.
|
868236 | if (weight_flag) |
581 | 203068 | fc->vvcdsp.inter.w_avg(dst, dst_stride, tmp[L0], tmp[L1], block_w, block_h, denom, w0, w1, o0, o1); | |
582 | else | ||
583 | 665168 | fc->vvcdsp.inter.avg(dst, dst_stride, tmp[L0], tmp[L1], block_w, block_h); | |
584 | 868236 | } | |
585 | |||
586 | 7429019 | static int pred_get_refs(const VVCLocalContext *lc, VVCRefPic *refp[2], const MvField *mv) | |
587 | { | ||
588 | 7429019 | RefPicList *rpl = lc->sc->rpl; | |
589 | |||
590 |
2/2✓ Branch 0 taken 14858038 times.
✓ Branch 1 taken 7429019 times.
|
22287057 | for (int mask = PF_L0; mask <= PF_L1; mask++) { |
591 |
2/2✓ Branch 0 taken 11174732 times.
✓ Branch 1 taken 3683306 times.
|
14858038 | if (mv->pred_flag & mask) { |
592 | 11174732 | const int lx = mask - PF_L0; | |
593 | 11174732 | refp[lx] = rpl[lx].refs + mv->ref_idx[lx]; | |
594 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 11174732 times.
|
11174732 | if (!refp[lx]->ref) |
595 | ✗ | return AVERROR_INVALIDDATA; | |
596 | } | ||
597 | } | ||
598 | 7429019 | return 0; | |
599 | } | ||
600 | |||
601 | #define POS(c_idx, x, y) \ | ||
602 | &fc->frame->data[c_idx][((y) >> fc->ps.sps->vshift[c_idx]) * fc->frame->linesize[c_idx] + \ | ||
603 | (((x) >> fc->ps.sps->hshift[c_idx]) << fc->ps.sps->pixel_shift)] | ||
604 | |||
605 | 24737 | static void pred_gpm_blk(VVCLocalContext *lc) | |
606 | { | ||
607 | 24737 | const VVCFrameContext *fc = lc->fc; | |
608 | 24737 | const CodingUnit *cu = lc->cu; | |
609 | 24737 | const PredictionUnit *pu = &cu->pu; | |
610 | |||
611 | 24737 | const uint8_t angle_idx = ff_vvc_gpm_angle_idx[pu->gpm_partition_idx]; | |
612 | 24737 | const uint8_t weights_idx = ff_vvc_gpm_angle_to_weights_idx[angle_idx]; | |
613 | 24737 | const int w = av_log2(cu->cb_width) - 3; | |
614 | 24737 | const int h = av_log2(cu->cb_height) - 3; | |
615 | 24737 | const uint8_t off_x = ff_vvc_gpm_weights_offset_x[pu->gpm_partition_idx][h][w]; | |
616 | 24737 | const uint8_t off_y = ff_vvc_gpm_weights_offset_y[pu->gpm_partition_idx][h][w]; | |
617 | 24737 | const uint8_t mirror_type = ff_vvc_gpm_angle_to_mirror[angle_idx]; | |
618 | const uint8_t *weights; | ||
619 | |||
620 |
2/2✓ Branch 0 taken 22456 times.
✓ Branch 1 taken 2281 times.
|
24737 | const int c_end = fc->ps.sps->r->sps_chroma_format_idc ? 3 : 1; |
621 | |||
622 | 24737 | int16_t *tmp[2] = {lc->tmp, lc->tmp1}; | |
623 | |||
624 |
2/2✓ Branch 0 taken 69649 times.
✓ Branch 1 taken 24737 times.
|
94386 | for (int c_idx = 0; c_idx < c_end; c_idx++) { |
625 | 69649 | const int hs = fc->ps.sps->hshift[c_idx]; | |
626 | 69649 | const int vs = fc->ps.sps->vshift[c_idx]; | |
627 | 69649 | const int x = lc->cu->x0 >> hs; | |
628 | 69649 | const int y = lc->cu->y0 >> vs; | |
629 | 69649 | const int width = cu->cb_width >> hs; | |
630 | 69649 | const int height = cu->cb_height >> vs; | |
631 | 69649 | uint8_t *dst = POS(c_idx, lc->cu->x0, lc->cu->y0); | |
632 | 69649 | ptrdiff_t dst_stride = fc->frame->linesize[c_idx]; | |
633 | |||
634 | 69649 | int step_x = 1 << hs; | |
635 | 69649 | int step_y = VVC_GPM_WEIGHT_SIZE << vs; | |
636 |
2/2✓ Branch 0 taken 43456 times.
✓ Branch 1 taken 26193 times.
|
69649 | if (!mirror_type) { |
637 | 43456 | weights = &ff_vvc_gpm_weights[weights_idx][off_y * VVC_GPM_WEIGHT_SIZE + off_x]; | |
638 |
2/2✓ Branch 0 taken 11195 times.
✓ Branch 1 taken 14998 times.
|
26193 | } else if (mirror_type == 1) { |
639 | 11195 | step_x = -step_x; | |
640 | 11195 | weights = &ff_vvc_gpm_weights[weights_idx][off_y * VVC_GPM_WEIGHT_SIZE + VVC_GPM_WEIGHT_SIZE - 1- off_x]; | |
641 | } else { | ||
642 | 14998 | step_y = -step_y; | |
643 | 14998 | weights = &ff_vvc_gpm_weights[weights_idx][(VVC_GPM_WEIGHT_SIZE - 1 - off_y) * VVC_GPM_WEIGHT_SIZE + off_x]; | |
644 | } | ||
645 | |||
646 |
2/2✓ Branch 0 taken 139298 times.
✓ Branch 1 taken 69649 times.
|
208947 | for (int i = 0; i < 2; i++) { |
647 | 139298 | const MvField *mv = pu->gpm_mv + i; | |
648 | 139298 | const int lx = mv->pred_flag - PF_L0; | |
649 | 139298 | VVCRefPic *refp = lc->sc->rpl[lx].refs + mv->ref_idx[lx]; | |
650 | |||
651 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 139298 times.
|
139298 | if (!refp->ref) |
652 | ✗ | return; | |
653 |
2/2✓ Branch 0 taken 819 times.
✓ Branch 1 taken 138479 times.
|
139298 | if (refp->is_scaled) |
654 | 819 | mc_scaled(lc, tmp[i], refp, mv->mv + lx, x, y, width, height, c_idx); | |
655 | else | ||
656 | 138479 | mc(lc, tmp[i], refp->ref, mv->mv + lx, x, y, width, height, c_idx); | |
657 | } | ||
658 | 69649 | fc->vvcdsp.inter.put_gpm(dst, dst_stride, width, height, tmp[0], tmp[1], weights, step_x, step_y); | |
659 | } | ||
660 | 24737 | return; | |
661 | } | ||
662 | |||
663 | 41475 | static int ciip_derive_intra_weight(const VVCLocalContext *lc, const int x0, const int y0, | |
664 | const int width, const int height) | ||
665 | { | ||
666 | 41475 | const VVCFrameContext *fc = lc->fc; | |
667 | 41475 | const VVCSPS *sps = fc->ps.sps; | |
668 | 41475 | const int x0b = av_zero_extend(x0, sps->ctb_log2_size_y); | |
669 | 41475 | const int y0b = av_zero_extend(y0, sps->ctb_log2_size_y); | |
670 |
4/4✓ Branch 0 taken 4546 times.
✓ Branch 1 taken 36929 times.
✓ Branch 2 taken 3698 times.
✓ Branch 3 taken 848 times.
|
41475 | const int available_l = lc->ctb_left_flag || x0b; |
671 |
4/4✓ Branch 0 taken 5970 times.
✓ Branch 1 taken 35505 times.
✓ Branch 2 taken 5211 times.
✓ Branch 3 taken 759 times.
|
41475 | const int available_u = lc->ctb_up_flag || y0b; |
672 | 41475 | const int min_pu_width = fc->ps.pps->min_pu_width; | |
673 | 41475 | int w = 1; | |
674 | |||
675 |
4/4✓ Branch 0 taken 40716 times.
✓ Branch 1 taken 759 times.
✓ Branch 2 taken 5893 times.
✓ Branch 3 taken 34823 times.
|
41475 | if (available_u &&fc->tab.mvf[((y0 - 1) >> MIN_PU_LOG2) * min_pu_width + ((x0 - 1 + width)>> MIN_PU_LOG2)].pred_flag == PF_INTRA) |
676 | 5893 | w++; | |
677 | |||
678 |
4/4✓ Branch 0 taken 40627 times.
✓ Branch 1 taken 848 times.
✓ Branch 2 taken 5621 times.
✓ Branch 3 taken 35006 times.
|
41475 | if (available_l && fc->tab.mvf[((y0 - 1 + height)>> MIN_PU_LOG2) * min_pu_width + ((x0 - 1) >> MIN_PU_LOG2)].pred_flag == PF_INTRA) |
679 | 5621 | w++; | |
680 | |||
681 | 41475 | return w; | |
682 | } | ||
683 | |||
684 | 3273413 | static void pred_regular(VVCLocalContext *lc, const MvField *mvf, const MvField *orig_mvf, | |
685 | const int x0, const int y0, const int sbw, const int sbh, const int sb_bdof_flag, const int c_start) | ||
686 | { | ||
687 | 3273413 | const VVCFrameContext *fc = lc->fc; | |
688 |
2/2✓ Branch 0 taken 3245126 times.
✓ Branch 1 taken 28287 times.
|
3273413 | const int c_end = fc->ps.sps->r->sps_chroma_format_idc ? CR : LUMA; |
689 | VVCRefPic *refp[2]; | ||
690 | |||
691 |
1/2✗ Branch 1 not taken.
✓ Branch 2 taken 3273413 times.
|
3273413 | if (pred_get_refs(lc, refp, mvf) < 0) |
692 | ✗ | return; | |
693 | |||
694 |
2/2✓ Branch 0 taken 8626567 times.
✓ Branch 1 taken 3273413 times.
|
11899980 | for (int c_idx = c_start; c_idx <= c_end; c_idx++) { |
695 | 8626567 | uint8_t *dst = POS(c_idx, x0, y0); | |
696 | 8626567 | const ptrdiff_t dst_stride = fc->frame->linesize[c_idx]; | |
697 | 8626567 | const int hs = fc->ps.sps->hshift[c_idx]; | |
698 | 8626567 | const int vs = fc->ps.sps->vshift[c_idx]; | |
699 | 8626567 | const int x = x0 >> hs; | |
700 | 8626567 | const int y = y0 >> vs; | |
701 | 8626567 | const int w = sbw >> hs; | |
702 | 8626567 | const int h = sbh >> vs; | |
703 | 8626567 | const int is_luma = !c_idx; | |
704 |
6/6✓ Branch 0 taken 45917 times.
✓ Branch 1 taken 8580650 times.
✓ Branch 2 taken 30400 times.
✓ Branch 3 taken 15517 times.
✓ Branch 4 taken 25958 times.
✓ Branch 5 taken 4442 times.
|
8626567 | const int do_ciip = lc->cu->ciip_flag && (is_luma || (w > 2)); |
705 |
2/2✓ Branch 0 taken 41475 times.
✓ Branch 1 taken 8585092 times.
|
8626567 | uint8_t *inter = do_ciip ? (uint8_t *)lc->ciip_tmp : dst; |
706 |
2/2✓ Branch 0 taken 8585092 times.
✓ Branch 1 taken 41475 times.
|
8626567 | const ptrdiff_t inter_stride = do_ciip ? (MAX_PB_SIZE * sizeof(uint16_t)) : dst_stride; |
707 |
4/4✓ Branch 0 taken 2136315 times.
✓ Branch 1 taken 6490252 times.
✓ Branch 2 taken 350143 times.
✓ Branch 3 taken 1786172 times.
|
8626567 | const int do_bdof = is_luma && sb_bdof_flag; |
708 | |||
709 |
2/2✓ Branch 0 taken 3355785 times.
✓ Branch 1 taken 5270782 times.
|
8626567 | if (mvf->pred_flag != PF_BI) { |
710 | 3355785 | const int lx = mvf->pred_flag - PF_L0; | |
711 | |||
712 |
2/2✓ Branch 0 taken 10749 times.
✓ Branch 1 taken 3345036 times.
|
3355785 | if (refp[lx]->is_scaled) { |
713 | 10749 | mc_uni_scaled(lc, inter, inter_stride, refp[lx], mvf, | |
714 | x, y, w, h, c_idx); | ||
715 | } else { | ||
716 | 3345036 | mc_uni(lc, inter, inter_stride, refp[lx]->ref, mvf, | |
717 | x, y, w, h, c_idx); | ||
718 | } | ||
719 | } else { | ||
720 |
4/4✓ Branch 0 taken 5269705 times.
✓ Branch 1 taken 1077 times.
✓ Branch 2 taken 9 times.
✓ Branch 3 taken 5269696 times.
|
5270782 | if (refp[L0]->is_scaled || refp[L1]->is_scaled) { |
721 | 1086 | mc_bi_scaled(lc, inter, inter_stride, refp[L0], refp[L1], mvf, | |
722 | x, y, w, h, c_idx); | ||
723 | } else { | ||
724 | 5269696 | mc_bi(lc, inter, inter_stride, refp[L0]->ref, refp[L1]->ref, mvf, orig_mvf, | |
725 | x, y, w, h, c_idx, do_bdof); | ||
726 | } | ||
727 | } | ||
728 |
2/2✓ Branch 0 taken 41475 times.
✓ Branch 1 taken 8585092 times.
|
8626567 | if (do_ciip) { |
729 | 41475 | const int intra_weight = ciip_derive_intra_weight(lc, x0, y0, sbw, sbh); | |
730 | 41475 | fc->vvcdsp.intra.intra_pred(lc, x0, y0, sbw, sbh, c_idx); | |
731 |
4/4✓ Branch 0 taken 15517 times.
✓ Branch 1 taken 25958 times.
✓ Branch 2 taken 8204 times.
✓ Branch 3 taken 7313 times.
|
41475 | if (!c_idx && lc->sc->sh.r->sh_lmcs_used_flag) |
732 | 8204 | fc->vvcdsp.lmcs.filter(inter, inter_stride, w, h, &fc->ps.lmcs.fwd_lut); | |
733 | 41475 | fc->vvcdsp.inter.put_ciip(dst, dst_stride, w, h, inter, inter_stride, intra_weight); | |
734 | } | ||
735 | } | ||
736 | } | ||
737 | |||
738 | // 8.5.3.5 Parametric motion vector refinement process | ||
739 | 495820 | static int parametric_mv_refine(const int *sad, const int stride) | |
740 | { | ||
741 | 495820 | const int sad_minus = sad[-stride]; | |
742 | 495820 | const int sad_center = sad[0]; | |
743 | 495820 | const int sad_plus = sad[stride]; | |
744 | int dmvc; | ||
745 | 495820 | int denom = (( sad_minus + sad_plus) - (sad_center << 1 ) ) << 3; | |
746 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 495820 times.
|
495820 | if (!denom) |
747 | ✗ | dmvc = 0; | |
748 | else { | ||
749 |
2/2✓ Branch 0 taken 58 times.
✓ Branch 1 taken 495762 times.
|
495820 | if (sad_minus == sad_center) |
750 | 58 | dmvc = -8; | |
751 |
2/2✓ Branch 0 taken 239 times.
✓ Branch 1 taken 495523 times.
|
495762 | else if (sad_plus == sad_center) |
752 | 239 | dmvc = 8; | |
753 | else { | ||
754 | 495523 | int num = ( sad_minus - sad_plus ) * (1 << 4); | |
755 | 495523 | int sign_num = 0; | |
756 | 495523 | int quotient = 0; | |
757 | 495523 | int counter = 3; | |
758 |
2/2✓ Branch 0 taken 250794 times.
✓ Branch 1 taken 244729 times.
|
495523 | if (num < 0 ) { |
759 | 250794 | num = - num; | |
760 | 250794 | sign_num = 1; | |
761 | } | ||
762 |
2/2✓ Branch 0 taken 1486569 times.
✓ Branch 1 taken 495523 times.
|
1982092 | while (counter > 0) { |
763 | 1486569 | counter = counter - 1; | |
764 | 1486569 | quotient = quotient << 1; | |
765 |
2/2✓ Branch 0 taken 440318 times.
✓ Branch 1 taken 1046251 times.
|
1486569 | if ( num >= denom ) { |
766 | 440318 | num = num - denom; | |
767 | 440318 | quotient = quotient + 1; | |
768 | } | ||
769 | 1486569 | denom = (denom >> 1); | |
770 | } | ||
771 |
2/2✓ Branch 0 taken 250794 times.
✓ Branch 1 taken 244729 times.
|
495523 | if (sign_num == 1 ) |
772 | 250794 | dmvc = -quotient; | |
773 | else | ||
774 | 244729 | dmvc = quotient; | |
775 | } | ||
776 | } | ||
777 | 495820 | return dmvc; | |
778 | } | ||
779 | |||
780 | #define SAD_ARRAY_SIZE 5 | ||
781 | //8.5.3 Decoder-side motion vector refinement process | ||
782 | 986366 | static void dmvr_mv_refine(VVCLocalContext *lc, MvField *mvf, MvField *orig_mv, int *sb_bdof_flag, | |
783 | const VVCFrame *ref0, const VVCFrame *ref1, const int x_off, const int y_off, const int block_w, const int block_h) | ||
784 | { | ||
785 | 986366 | const VVCFrameContext *fc = lc->fc; | |
786 | 986366 | const int sr_range = 2; | |
787 | 986366 | const VVCFrame *refs[] = { ref0, ref1 }; | |
788 | 986366 | int16_t *tmp[] = { lc->tmp, lc->tmp1 }; | |
789 | int sad[SAD_ARRAY_SIZE][SAD_ARRAY_SIZE]; | ||
790 | int min_dx, min_dy, min_sad, dx, dy; | ||
791 | |||
792 | 986366 | *orig_mv = *mvf; | |
793 | 986366 | min_dx = min_dy = dx = dy = 2; | |
794 | |||
795 |
2/2✓ Branch 0 taken 1972732 times.
✓ Branch 1 taken 986366 times.
|
2959098 | for (int i = L0; i <= L1; i++) { |
796 | 1972732 | const int pred_w = block_w + 2 * sr_range; | |
797 | 1972732 | const int pred_h = block_h + 2 * sr_range; | |
798 | 1972732 | const Mv *mv = mvf->mv + i; | |
799 | 1972732 | const int mx = mv->x & 0xf; | |
800 | 1972732 | const int my = mv->y & 0xf; | |
801 | 1972732 | const int ox = x_off + (mv->x >> 4) - sr_range; | |
802 | 1972732 | const int oy = y_off + (mv->y >> 4) - sr_range; | |
803 | 1972732 | const VVCFrame *ref = refs[i]; | |
804 | 1972732 | ptrdiff_t src_stride = ref->frame->linesize[LUMA]; | |
805 | 1972732 | const uint8_t *src = ref->frame->data[LUMA]; | |
806 | 1972732 | const int wrap_enabled = fc->ps.pps->r->pps_ref_wraparound_enabled_flag; | |
807 | |||
808 | 1972732 | MC_EMULATED_EDGE_BILINEAR(lc->edge_emu_buffer, &src, &src_stride, ox, oy); | |
809 | 1972732 | fc->vvcdsp.inter.dmvr[!!my][!!mx](tmp[i], src, src_stride, pred_h, mx, my, pred_w); | |
810 | } | ||
811 | |||
812 | 986366 | min_sad = fc->vvcdsp.inter.sad(tmp[L0], tmp[L1], dx, dy, block_w, block_h); | |
813 | 986366 | min_sad -= min_sad >> 2; | |
814 | 986366 | sad[dy][dx] = min_sad; | |
815 | |||
816 |
2/2✓ Branch 0 taken 449926 times.
✓ Branch 1 taken 536440 times.
|
986366 | if (min_sad >= block_w * block_h) { |
817 | int dmv[2]; | ||
818 | // 8.5.3.4 Array entry selection process | ||
819 |
2/2✓ Branch 0 taken 2249630 times.
✓ Branch 1 taken 449926 times.
|
2699556 | for (dy = 0; dy < SAD_ARRAY_SIZE; dy++) { |
820 |
2/2✓ Branch 0 taken 11248150 times.
✓ Branch 1 taken 2249630 times.
|
13497780 | for (dx = 0; dx < SAD_ARRAY_SIZE; dx++) { |
821 |
4/4✓ Branch 0 taken 2249630 times.
✓ Branch 1 taken 8998520 times.
✓ Branch 2 taken 1799704 times.
✓ Branch 3 taken 449926 times.
|
11248150 | if (dx != sr_range || dy != sr_range) { |
822 | 10798224 | sad[dy][dx] = fc->vvcdsp.inter.sad(lc->tmp, lc->tmp1, dx, dy, block_w, block_h); | |
823 |
2/2✓ Branch 0 taken 645265 times.
✓ Branch 1 taken 10152959 times.
|
10798224 | if (sad[dy][dx] < min_sad) { |
824 | 645265 | min_sad = sad[dy][dx]; | |
825 | 645265 | min_dx = dx; | |
826 | 645265 | min_dy = dy; | |
827 | } | ||
828 | } | ||
829 | } | ||
830 | } | ||
831 | 449926 | dmv[0] = (min_dx - sr_range) * (1 << 4); | |
832 | 449926 | dmv[1] = (min_dy - sr_range) * (1 << 4); | |
833 |
8/8✓ Branch 0 taken 372784 times.
✓ Branch 1 taken 77142 times.
✓ Branch 2 taken 308341 times.
✓ Branch 3 taken 64443 times.
✓ Branch 4 taken 256333 times.
✓ Branch 5 taken 52008 times.
✓ Branch 6 taken 247910 times.
✓ Branch 7 taken 8423 times.
|
449926 | if (min_dx != 0 && min_dx != 4 && min_dy != 0 && min_dy != 4) { |
834 | 247910 | dmv[0] += parametric_mv_refine(&sad[min_dy][min_dx], 1); | |
835 | 247910 | dmv[1] += parametric_mv_refine(&sad[min_dy][min_dx], SAD_ARRAY_SIZE); | |
836 | } | ||
837 | |||
838 |
2/2✓ Branch 0 taken 899852 times.
✓ Branch 1 taken 449926 times.
|
1349778 | for (int i = L0; i <= L1; i++) { |
839 | 899852 | Mv *mv = mvf->mv + i; | |
840 | 899852 | mv->x += (1 - 2 * i) * dmv[0]; | |
841 | 899852 | mv->y += (1 - 2 * i) * dmv[1]; | |
842 | 899852 | ff_vvc_clip_mv(mv); | |
843 | } | ||
844 | } | ||
845 |
2/2✓ Branch 0 taken 701735 times.
✓ Branch 1 taken 284631 times.
|
986366 | if (min_sad < 2 * block_w * block_h) { |
846 | 701735 | *sb_bdof_flag = 0; | |
847 | } | ||
848 | 986366 | } | |
849 | |||
850 | 986366 | static void set_dmvr_info(VVCFrameContext *fc, const int x0, const int y0, | |
851 | const int width, const int height, const MvField *mvf) | ||
852 | |||
853 | { | ||
854 | 986366 | const VVCPPS *pps = fc->ps.pps; | |
855 | |||
856 |
2/2✓ Branch 0 taken 3903492 times.
✓ Branch 1 taken 986366 times.
|
4889858 | for (int y = y0; y < y0 + height; y += MIN_PU_SIZE) { |
857 |
2/2✓ Branch 0 taken 15480184 times.
✓ Branch 1 taken 3903492 times.
|
19383676 | for (int x = x0; x < x0 + width; x += MIN_PU_SIZE) { |
858 | 15480184 | const int idx = pps->min_pu_width * (y >> MIN_PU_LOG2) + (x >> MIN_PU_LOG2); | |
859 | 15480184 | fc->ref->tab_dmvr_mvf[idx] = *mvf; | |
860 | } | ||
861 | } | ||
862 | 986366 | } | |
863 | |||
864 | 2136315 | static void derive_sb_mv(VVCLocalContext *lc, MvField *mv, MvField *orig_mv, int *sb_bdof_flag, | |
865 | const int x0, const int y0, const int sbw, const int sbh) | ||
866 | { | ||
867 | 2136315 | VVCFrameContext *fc = lc->fc; | |
868 | 2136315 | const PredictionUnit *pu = &lc->cu->pu; | |
869 | |||
870 | 2136315 | *orig_mv = *mv = *ff_vvc_get_mvf(fc, x0, y0); | |
871 |
2/2✓ Branch 0 taken 1051878 times.
✓ Branch 1 taken 1084437 times.
|
2136315 | if (pu->bdof_flag) |
872 | 1051878 | *sb_bdof_flag = 1; | |
873 |
2/2✓ Branch 0 taken 986366 times.
✓ Branch 1 taken 1149949 times.
|
2136315 | if (pu->dmvr_flag) { |
874 | VVCRefPic *refp[2]; | ||
875 |
1/2✗ Branch 1 not taken.
✓ Branch 2 taken 986366 times.
|
986366 | if (pred_get_refs(lc, refp, mv) < 0) |
876 | ✗ | return; | |
877 | 986366 | dmvr_mv_refine(lc, mv, orig_mv, sb_bdof_flag, refp[L0]->ref, refp[L1]->ref, x0, y0, sbw, sbh); | |
878 | 986366 | set_dmvr_info(fc, x0, y0, sbw, sbh, mv); | |
879 | } | ||
880 | } | ||
881 | |||
882 | 362179 | static void pred_regular_blk(VVCLocalContext *lc, const int skip_ciip) | |
883 | { | ||
884 | 362179 | const CodingUnit *cu = lc->cu; | |
885 | 362179 | PredictionUnit *pu = &lc->cu->pu; | |
886 | 362179 | const MotionInfo *mi = &pu->mi; | |
887 | MvField mv, orig_mv; | ||
888 | 362179 | int sbw, sbh, sb_bdof_flag = 0; | |
889 | |||
890 |
4/4✓ Branch 0 taken 31034 times.
✓ Branch 1 taken 331145 times.
✓ Branch 2 taken 15517 times.
✓ Branch 3 taken 15517 times.
|
362179 | if (cu->ciip_flag && skip_ciip) |
891 | 15517 | return; | |
892 | |||
893 | 346662 | sbw = cu->cb_width / mi->num_sb_x; | |
894 | 346662 | sbh = cu->cb_height / mi->num_sb_y; | |
895 | |||
896 |
2/2✓ Branch 0 taken 572896 times.
✓ Branch 1 taken 346662 times.
|
919558 | for (int sby = 0; sby < mi->num_sb_y; sby++) { |
897 |
2/2✓ Branch 0 taken 2136315 times.
✓ Branch 1 taken 572896 times.
|
2709211 | for (int sbx = 0; sbx < mi->num_sb_x; sbx++) { |
898 | 2136315 | const int x0 = cu->x0 + sbx * sbw; | |
899 | 2136315 | const int y0 = cu->y0 + sby * sbh; | |
900 | |||
901 |
2/2✓ Branch 0 taken 15517 times.
✓ Branch 1 taken 2120798 times.
|
2136315 | if (cu->ciip_flag) |
902 | 15517 | ff_vvc_set_neighbour_available(lc, x0, y0, sbw, sbh); | |
903 | |||
904 | 2136315 | derive_sb_mv(lc, &mv, &orig_mv, &sb_bdof_flag, x0, y0, sbw, sbh); | |
905 | 2136315 | pred_regular(lc, &mv, &orig_mv, x0, y0, sbw, sbh, sb_bdof_flag, LUMA); | |
906 | } | ||
907 | } | ||
908 | } | ||
909 | |||
910 | 1137098 | static void derive_affine_mvc(MvField *mvc, const VVCFrameContext *fc, const MvField *mv, | |
911 | const int x0, const int y0, const int sbw, const int sbh) | ||
912 | { | ||
913 | 1137098 | const int hs = fc->ps.sps->hshift[1]; | |
914 | 1137098 | const int vs = fc->ps.sps->vshift[1]; | |
915 | 1137098 | const MvField* mv2 = ff_vvc_get_mvf(fc, x0 + hs * sbw, y0 + vs * sbh); | |
916 | 1137098 | *mvc = *mv; | |
917 | |||
918 | // Due to different pred_flag, one of the motion vectors may have an invalid value. | ||
919 | // Cast them to an unsigned type to avoid undefined behavior. | ||
920 | 1137098 | mvc->mv[0].x += (unsigned int)mv2->mv[0].x; | |
921 | 1137098 | mvc->mv[0].y += (unsigned int)mv2->mv[0].y; | |
922 | 1137098 | mvc->mv[1].x += (unsigned int)mv2->mv[1].x; | |
923 | 1137098 | mvc->mv[1].y += (unsigned int)mv2->mv[1].y; | |
924 | 1137098 | ff_vvc_round_mv(mvc->mv + 0, 0, 1); | |
925 | 1137098 | ff_vvc_round_mv(mvc->mv + 1, 0, 1); | |
926 | 1137098 | } | |
927 | |||
928 | 32577 | static void pred_affine_blk(VVCLocalContext *lc) | |
929 | { | ||
930 | 32577 | const VVCFrameContext *fc = lc->fc; | |
931 | 32577 | const CodingUnit *cu = lc->cu; | |
932 | 32577 | const PredictionUnit *pu = &cu->pu; | |
933 | 32577 | const MotionInfo *mi = &pu->mi; | |
934 | 32577 | const int x0 = cu->x0; | |
935 | 32577 | const int y0 = cu->y0; | |
936 | 32577 | const int sbw = cu->cb_width / mi->num_sb_x; | |
937 | 32577 | const int sbh = cu->cb_height / mi->num_sb_y; | |
938 | 32577 | const int hs = fc->ps.sps->hshift[1]; | |
939 | 32577 | const int vs = fc->ps.sps->vshift[1]; | |
940 | 32577 | const int dst_stride = fc->frame->linesize[LUMA]; | |
941 | |||
942 |
2/2✓ Branch 0 taken 258140 times.
✓ Branch 1 taken 32577 times.
|
290717 | for (int sby = 0; sby < mi->num_sb_y; sby++) { |
943 |
2/2✓ Branch 0 taken 3169240 times.
✓ Branch 1 taken 258140 times.
|
3427380 | for (int sbx = 0; sbx < mi->num_sb_x; sbx++) { |
944 | 3169240 | const int x = x0 + sbx * sbw; | |
945 | 3169240 | const int y = y0 + sby * sbh; | |
946 | |||
947 | 3169240 | uint8_t *dst0 = POS(0, x, y); | |
948 | 3169240 | const MvField *mv = ff_vvc_get_mvf(fc, x, y); | |
949 | VVCRefPic *refp[2]; | ||
950 | |||
951 |
1/2✗ Branch 1 not taken.
✓ Branch 2 taken 3169240 times.
|
3169240 | if (pred_get_refs(lc, refp, mv) < 0) |
952 | ✗ | return; | |
953 | |||
954 |
2/2✓ Branch 0 taken 2301004 times.
✓ Branch 1 taken 868236 times.
|
3169240 | if (mi->pred_flag != PF_BI) { |
955 | 2301004 | const int lx = mi->pred_flag - PF_L0; | |
956 |
2/2✓ Branch 0 taken 16008 times.
✓ Branch 1 taken 2284996 times.
|
2301004 | if (refp[lx]->is_scaled) { |
957 | 16008 | mc_uni_scaled(lc, dst0, dst_stride, refp[lx], mv, x, y, sbw, sbh, LUMA); | |
958 | } else { | ||
959 | 2284996 | luma_prof_uni(lc, dst0, dst_stride, refp[lx]->ref, | |
960 | 2284996 | mv, x, y, sbw, sbh, pu->cb_prof_flag[lx], | |
961 | 2284996 | pu->diff_mv_x[lx], pu->diff_mv_y[lx]); | |
962 | } | ||
963 | } else { | ||
964 | 868236 | luma_prof_bi(lc, dst0, dst_stride, refp[L0], refp[L1], mv, x, y, sbw, sbh); | |
965 | } | ||
966 |
2/2✓ Branch 0 taken 3131132 times.
✓ Branch 1 taken 38108 times.
|
3169240 | if (fc->ps.sps->r->sps_chroma_format_idc) { |
967 |
4/4✓ Branch 0 taken 1801776 times.
✓ Branch 1 taken 1329356 times.
✓ Branch 2 taken 1137098 times.
✓ Branch 3 taken 664678 times.
|
3131132 | if (!av_zero_extend(sby, vs) && !av_zero_extend(sbx, hs)) { |
968 | MvField mvc; | ||
969 | |||
970 | 1137098 | derive_affine_mvc(&mvc, fc, mv, x, y, sbw, sbh); | |
971 | 1137098 | pred_regular(lc, &mvc, NULL, x, y, sbw << hs, sbh << vs, 0, CB); | |
972 | } | ||
973 | } | ||
974 | |||
975 | } | ||
976 | } | ||
977 | } | ||
978 | |||
979 | 403976 | static void predict_inter(VVCLocalContext *lc) | |
980 | { | ||
981 | 403976 | const VVCFrameContext *fc = lc->fc; | |
982 | 403976 | const CodingUnit *cu = lc->cu; | |
983 | 403976 | const PredictionUnit *pu = &cu->pu; | |
984 | |||
985 |
2/2✓ Branch 0 taken 24737 times.
✓ Branch 1 taken 379239 times.
|
403976 | if (pu->merge_gpm_flag) |
986 | 24737 | pred_gpm_blk(lc); | |
987 |
2/2✓ Branch 0 taken 32577 times.
✓ Branch 1 taken 346662 times.
|
379239 | else if (pu->inter_affine_flag) |
988 | 32577 | pred_affine_blk(lc); | |
989 | else | ||
990 | 346662 | pred_regular_blk(lc, 1); //intra block is not ready yet, skip ciip | |
991 | |||
992 |
4/4✓ Branch 0 taken 174368 times.
✓ Branch 1 taken 229608 times.
✓ Branch 2 taken 166164 times.
✓ Branch 3 taken 8204 times.
|
403976 | if (lc->sc->sh.r->sh_lmcs_used_flag && !cu->ciip_flag) { |
993 | 166164 | uint8_t* dst0 = POS(0, cu->x0, cu->y0); | |
994 | 166164 | fc->vvcdsp.lmcs.filter(dst0, fc->frame->linesize[LUMA], cu->cb_width, cu->cb_height, &fc->ps.lmcs.fwd_lut); | |
995 | } | ||
996 | 403976 | } | |
997 | |||
998 | 1107877 | static int has_inter_luma(const CodingUnit *cu) | |
999 | { | ||
1000 |
4/6✓ Branch 0 taken 703901 times.
✓ Branch 1 taken 403976 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 703901 times.
✓ Branch 4 taken 403976 times.
✗ Branch 5 not taken.
|
1107877 | return (cu->pred_mode == MODE_INTER || cu->pred_mode == MODE_SKIP) && cu->tree_type != DUAL_TREE_CHROMA; |
1001 | } | ||
1002 | |||
1003 | 45792 | int ff_vvc_predict_inter(VVCLocalContext *lc, const int rs) | |
1004 | { | ||
1005 | 45792 | const VVCFrameContext *fc = lc->fc; | |
1006 | 45792 | CodingUnit *cu = fc->tab.cus[rs]; | |
1007 | |||
1008 |
2/2✓ Branch 0 taken 1107877 times.
✓ Branch 1 taken 45792 times.
|
1153669 | while (cu) { |
1009 | 1107877 | lc->cu = cu; | |
1010 |
2/2✓ Branch 1 taken 403976 times.
✓ Branch 2 taken 703901 times.
|
1107877 | if (has_inter_luma(cu)) |
1011 | 403976 | predict_inter(lc); | |
1012 | 1107877 | cu = cu->next; | |
1013 | } | ||
1014 | |||
1015 | 45792 | return 0; | |
1016 | } | ||
1017 | |||
1018 | 15517 | void ff_vvc_predict_ciip(VVCLocalContext *lc) | |
1019 | { | ||
1020 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 15517 times.
|
15517 | av_assert0(lc->cu->ciip_flag); |
1021 | |||
1022 | //todo: refact out ciip from pred_regular_blk | ||
1023 | 15517 | pred_regular_blk(lc, 0); | |
1024 | 15517 | } | |
1025 | |||
1026 | #undef POS | ||
1027 |