History

vvcdec: add motion vector decoder

Nuo Mi committed 2 years ago

Commit 603d0bd1 parent c1a3d175 files · +1865 −0Browse files

Unified · Split

5 files
  1. libavcodec/vvc/Makefile +1 −0
  2. libavcodec/vvc/vvc_ctu.c +18 −0
  3. libavcodec/vvc/vvc_ctu.h +2 −0
  4. libavcodec/vvc/vvc_mvs.c +1798 −0
  5. libavcodec/vvc/vvc_mvs.h +46 −0
libavcodec/vvc/Makefile+1−0
@@ -4,5 +4,6 @@ clean:: · 3 unmodified lines
44 OBJS-$(CONFIG_VVC_DECODER) += vvc/vvc_cabac.o \
55 vvc/vvc_ctu.o \
66 vvc/vvc_data.o \
7+ vvc/vvc_mvs.o \
78 vvc/vvc_ps.o \
89 vvc/vvc_refs.o \
libavcodec/vvc/vvc_ctu.c+18−0
@@ -20,7 +20,25 @@ · 19 unmodified lines
2020 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
2121 */
2222
23+#include "vvc_cabac.h"
2324 #include "vvc_ctu.h"
25+#include "vvc_mvs.h"
26+
27+void ff_vvc_set_neighbour_available(VVCLocalContext *lc,
28+ const int x0, const int y0, const int w, const int h)
29+{
30+ const int log2_ctb_size = lc->fc->ps.sps->ctb_log2_size_y;
31+ const int x0b = av_mod_uintp2(x0, log2_ctb_size);
32+ const int y0b = av_mod_uintp2(y0, log2_ctb_size);
33+
34+ lc->na.cand_up = (lc->ctb_up_flag || y0b);
35+ lc->na.cand_left = (lc->ctb_left_flag || x0b);
36+ lc->na.cand_up_left = (x0b || y0b) ? lc->na.cand_left && lc->na.cand_up : lc->ctb_up_left_flag;
37+ lc->na.cand_up_right_sap =
38+ (x0b + w == 1 << log2_ctb_size) ? lc->ctb_up_right_flag && !y0b : lc->na.cand_up;
39+ lc->na.cand_up_right = lc->na.cand_up_right_sap && (x0 + w) < lc->end_of_tiles_x;
40+}
41+
2442
2543 void ff_vvc_ep_init_stat_coeff(EntryPoint *ep,
2644 const int bit_depth, const int persistent_rice_adaptation_enabled_flag)
libavcodec/vvc/vvc_ctu.h+2−0
@@ -459,6 +459,8 @@ typedef struct ALFParams { · 458 unmodified lines
459459 uint8_t applied[3];
460460 } ALFParams;
461461
462+//utils
463+void ff_vvc_set_neighbour_available(VVCLocalContext *lc, int x0, int y0, int w, int h);
462464 void ff_vvc_ep_init_stat_coeff(EntryPoint *ep, int bit_depth, int persistent_rice_adaptation_enabled_flag);
463465
464466 #endif // AVCODEC_VVC_VVC_CTU_H
libavcodec/vvc/vvc_mvs.c+1798−0
@@ -0,0 +1,1798 @@
1+/*
2+ * VVC motion vector decoder
3+ *
4+ * Copyright (C) 2023 Nuo Mi
5+ * Copyright (C) 2022 Xu Mu
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+
23+#include "vvc_ctu.h"
24+#include "vvc_data.h"
25+#include "vvc_refs.h"
26+#include "vvc_mvs.h"
27+
28+#define IS_SAME_MV(a, b) (AV_RN64A(a) == AV_RN64A(b))
29+
30+//check if the two luma locations belong to the same motion estimation region
31+static av_always_inline int is_same_mer(const VVCFrameContext *fc, const int xN, const int yN, const int xP, const int yP)
32+{
33+ const uint8_t plevel = fc->ps.sps->log2_parallel_merge_level;
34+
35+ return xN >> plevel == xP >> plevel &&
36+ yN >> plevel == yP >> plevel;
37+}
38+
39+//return true if we have same mvs and ref_idxs
40+static av_always_inline int compare_mv_ref_idx(const MvField *n, const MvField *o)
41+{
42+ if (!o || n->pred_flag != o->pred_flag)
43+ return 0;
44+ for (int i = 0; i < 2; i++) {
45+ PredFlag mask = i + 1;
46+ if (n->pred_flag & mask) {
47+ const int same_ref_idx = n->ref_idx[i] == o->ref_idx[i];
48+ const int same_mv = IS_SAME_MV(n->mv + i, o->mv + i);
49+ if (!same_ref_idx || !same_mv)
50+ return 0;
51+ }
52+ }
53+ return 1;
54+}
55+
56+// 8.5.2.15 Temporal motion buffer compression process for collocated motion vectors
57+static av_always_inline void mv_compression(Mv *motion)
58+{
59+ int mv[2] = {motion->x, motion->y};
60+ for (int i = 0; i < 2; i++) {
61+ const int s = mv[i] >> 17;
62+ const int f = av_log2((mv[i] ^ s) | 31) - 4;
63+ const int mask = (-1 << f) >> 1;
64+ const int round = (1 << f) >> 2;
65+ mv[i] = (mv[i] + round) & mask;
66+ }
67+ motion->x = mv[0];
68+ motion->y = mv[1];
69+}
70+
71+void ff_vvc_mv_scale(Mv *dst, const Mv *src, int td, int tb)
72+{
73+ int tx, scale_factor;
74+
75+ td = av_clip_int8(td);
76+ tb = av_clip_int8(tb);
77+ tx = (0x4000 + (abs(td) >> 1)) / td;
78+ scale_factor = av_clip_intp2((tb * tx + 32) >> 6, 12);
79+ dst->x = av_clip_intp2((scale_factor * src->x + 127 +
80+ (scale_factor * src->x < 0)) >> 8, 17);
81+ dst->y = av_clip_intp2((scale_factor * src->y + 127 +
82+ (scale_factor * src->y < 0)) >> 8, 17);
83+}
84+
85+//part of 8.5.2.12 Derivation process for collocated motion vectors
86+static int check_mvset(Mv *mvLXCol, Mv *mvCol,
87+ int colPic, int poc,
88+ const RefPicList *refPicList, int X, int refIdxLx,
89+ const RefPicList *refPicList_col, int listCol, int refidxCol)
90+{
91+ int cur_lt = refPicList[X].isLongTerm[refIdxLx];
92+ int col_lt = refPicList_col[listCol].isLongTerm[refidxCol];
93+ int col_poc_diff, cur_poc_diff;
94+
95+ if (cur_lt != col_lt) {
96+ mvLXCol->x = 0;
97+ mvLXCol->y = 0;
98+ return 0;
99+ }
100+
101+ col_poc_diff = colPic - refPicList_col[listCol].list[refidxCol];
102+ cur_poc_diff = poc - refPicList[X].list[refIdxLx];
103+
104+ mv_compression(mvCol);
105+ if (cur_lt || col_poc_diff == cur_poc_diff) {
106+ mvLXCol->x = av_clip_intp2(mvCol->x, 17);
107+ mvLXCol->y = av_clip_intp2(mvCol->y, 17);
108+ } else {
109+ ff_vvc_mv_scale(mvLXCol, mvCol, col_poc_diff, cur_poc_diff);
110+ }
111+ return 1;
112+}
113+
114+#define CHECK_MVSET(l) \
115+ check_mvset(mvLXCol, temp_col.mv + l, \
116+ colPic, fc->ps.ph.poc, \
117+ refPicList, X, refIdxLx, \
118+ refPicList_col, L ## l, temp_col.ref_idx[l])
119+
120+//derive NoBackwardPredFlag
121+int ff_vvc_no_backward_pred_flag(const VVCLocalContext *lc)
122+{
123+ int check_diffpicount = 0;
124+ int i, j;
125+ const RefPicList *rpl = lc->sc->rpl;
126+
127+ for (j = 0; j < 2; j++) {
128+ for (i = 0; i < rpl[j].nb_refs; i++) {
129+ if (rpl[j].list[i] > lc->fc->ps.ph.poc) {
130+ check_diffpicount++;
131+ break;
132+ }
133+ }
134+ }
135+ return !check_diffpicount;
136+}
137+
138+//8.5.2.12 Derivation process for collocated motion vectors
139+static int derive_temporal_colocated_mvs(const VVCLocalContext *lc, MvField temp_col,
140+ int refIdxLx, Mv *mvLXCol, int X,
141+ int colPic, const RefPicList *refPicList_col, int sb_flag)
142+{
143+ const VVCFrameContext *fc = lc->fc;
144+ const SliceContext *sc = lc->sc;
145+ RefPicList* refPicList = sc->rpl;
146+
147+ if (temp_col.pred_flag == PF_INTRA)
148+ return 0;
149+
150+ if (sb_flag){
151+ if (X == 0) {
152+ if (temp_col.pred_flag & PF_L0)
153+ return CHECK_MVSET(0);
154+ else if (ff_vvc_no_backward_pred_flag(lc) && (temp_col.pred_flag & PF_L1))
155+ return CHECK_MVSET(1);
156+ } else {
157+ if (temp_col.pred_flag & PF_L1)
158+ return CHECK_MVSET(1);
159+ else if (ff_vvc_no_backward_pred_flag(lc) && (temp_col.pred_flag & PF_L0))
160+ return CHECK_MVSET(0);
161+ }
162+ } else {
163+ if (!(temp_col.pred_flag & PF_L0))
164+ return CHECK_MVSET(1);
165+ else if (temp_col.pred_flag == PF_L0)
166+ return CHECK_MVSET(0);
167+ else if (temp_col.pred_flag == PF_BI) {
168+ if (ff_vvc_no_backward_pred_flag(lc)) {
169+ if (X == 0)
170+ return CHECK_MVSET(0);
171+ else
172+ return CHECK_MVSET(1);
173+ } else {
174+ if (!lc->sc->sh.r->sh_collocated_from_l0_flag)
175+ return CHECK_MVSET(0);
176+ else
177+ return CHECK_MVSET(1);
178+ }
179+ }
180+ }
181+ return 0;
182+}
183+
184+#define TAB_MVF(x, y) \
185+ tab_mvf[((y) >> MIN_PU_LOG2) * min_pu_width + ((x) >> MIN_PU_LOG2)]
186+
187+#define TAB_MVF_PU(v) \
188+ TAB_MVF(x ## v, y ## v)
189+
190+#define TAB_CP_MV(lx, x, y) \
191+ fc->tab.cp_mv[lx][((((y) >> min_cb_log2_size) * min_cb_width + ((x) >> min_cb_log2_size)) ) * MAX_CONTROL_POINTS]
192+
193+
194+#define DERIVE_TEMPORAL_COLOCATED_MVS(sb_flag) \
195+ derive_temporal_colocated_mvs(lc, temp_col, \
196+ refIdxLx, mvLXCol, X, colPic, \
197+ ff_vvc_get_ref_list(fc, ref, x, y), sb_flag)
198+
199+//8.5.2.11 Derivation process for temporal luma motion vector prediction
200+static int temporal_luma_motion_vector(const VVCLocalContext *lc,
201+ const int refIdxLx, Mv *mvLXCol, const int X, int check_center, int sb_flag)
202+{
203+ const VVCFrameContext *fc = lc->fc;
204+ const VVCSPS *sps = fc->ps.sps;
205+ const CodingUnit *cu = lc->cu;
206+ int x, y, colPic, availableFlagLXCol = 0;
207+ int min_pu_width = fc->ps.pps->min_pu_width;
208+ VVCFrame *ref = fc->ref->collocated_ref;
209+ MvField *tab_mvf;
210+ MvField temp_col;
211+
212+ if (!ref) {
213+ memset(mvLXCol, 0, sizeof(*mvLXCol));
214+ return 0;
215+ }
216+
217+ if (!fc->ps.ph.r->ph_temporal_mvp_enabled_flag || (cu->cb_width * cu->cb_height <= 32))
218+ return 0;
219+
220+ tab_mvf = ref->tab_dmvr_mvf;
221+ colPic = ref->poc;
222+
223+ //bottom right collocated motion vector
224+ x = cu->x0 + cu->cb_width;
225+ y = cu->y0 + cu->cb_height;
226+
227+ if (tab_mvf &&
228+ (cu->y0 >> sps->ctb_log2_size_y) == (y >> sps->ctb_log2_size_y) &&
229+ y < fc->ps.sps->height &&
230+ x < fc->ps.sps->width) {
231+ x &= ~7;
232+ y &= ~7;
233+ temp_col = TAB_MVF(x, y);
234+ availableFlagLXCol = DERIVE_TEMPORAL_COLOCATED_MVS(sb_flag);
235+ }
236+ if (check_center) {
237+ // derive center collocated motion vector
238+ if (tab_mvf && !availableFlagLXCol) {
239+ x = cu->x0 + (cu->cb_width >> 1);
240+ y = cu->y0 + (cu->cb_height >> 1);
241+ x &= ~7;
242+ y &= ~7;
243+ temp_col = TAB_MVF(x, y);
244+ availableFlagLXCol = DERIVE_TEMPORAL_COLOCATED_MVS(sb_flag);
245+ }
246+ }
247+ return availableFlagLXCol;
248+}
249+
250+void ff_vvc_set_mvf(const VVCLocalContext *lc, const int x0, const int y0, const int w, const int h, const MvField *mvf)
251+{
252+ const VVCFrameContext *fc = lc->fc;
253+ MvField *tab_mvf = fc->tab.mvf;
254+ const int min_pu_width = fc->ps.pps->min_pu_width;
255+ const int min_pu_size = 1 << MIN_PU_LOG2;
256+ for (int dy = 0; dy < h; dy += min_pu_size) {
257+ for (int dx = 0; dx < w; dx += min_pu_size) {
258+ const int x = x0 + dx;
259+ const int y = y0 + dy;
260+ TAB_MVF(x, y) = *mvf;
261+ }
262+ }
263+}
264+
265+void ff_vvc_set_intra_mvf(const VVCLocalContext *lc)
266+{
267+ const VVCFrameContext *fc = lc->fc;
268+ const CodingUnit *cu = lc->cu;
269+ MvField *tab_mvf = fc->tab.mvf;
270+ const int min_pu_width = fc->ps.pps->min_pu_width;
271+ const int min_pu_size = 1 << MIN_PU_LOG2;
272+ for (int dy = 0; dy < cu->cb_height; dy += min_pu_size) {
273+ for (int dx = 0; dx < cu->cb_width; dx += min_pu_size) {
274+ const int x = cu->x0 + dx;
275+ const int y = cu->y0 + dy;
276+ TAB_MVF(x, y).pred_flag = PF_INTRA;
277+ }
278+ }
279+}
280+
281+//cbProfFlagLX from 8.5.5.9 Derivation process for motion vector arrays from affine control point motion vectors
282+static int derive_cb_prof_flag_lx(const VVCLocalContext *lc, const PredictionUnit* pu, int lx, int is_fallback)
283+{
284+ const MotionInfo* mi = &pu->mi;
285+ const Mv* cp_mv = &mi->mv[lx][0];
286+ if (lc->fc->ps.ph.r->ph_prof_disabled_flag || is_fallback)
287+ return 0;
288+ if (mi->motion_model_idc == MOTION_4_PARAMS_AFFINE) {
289+ if (IS_SAME_MV(cp_mv, cp_mv + 1))
290+ return 0;
291+ }
292+ if (mi->motion_model_idc == MOTION_6_PARAMS_AFFINE) {
293+ if (IS_SAME_MV(cp_mv, cp_mv + 1) && IS_SAME_MV(cp_mv, cp_mv + 2))
294+ return 0;
295+ }
296+ //fixme: RprConstraintsActiveFlag
297+ return 1;
298+}
299+
300+typedef struct SubblockParams {
301+ int d_hor_x;
302+ int d_ver_x;
303+ int d_hor_y;
304+ int d_ver_y;
305+ int mv_scale_hor;
306+ int mv_scale_ver;
307+ int is_fallback;
308+
309+ int cb_width;
310+ int cb_height;
311+} SubblockParams;
312+
313+static int is_fallback_mode(const SubblockParams *sp, const PredFlag pred_flag)
314+{
315+ const int a = 4 * (2048 + sp->d_hor_x);
316+ const int b = 4 * sp->d_hor_y;
317+ const int c = 4 * (2048 + sp->d_ver_y);
318+ const int d = 4 * sp->d_ver_x;
319+ if (pred_flag == PF_BI) {
320+ const int max_w4 = FFMAX(0, FFMAX(a, FFMAX(b, a + b)));
321+ const int min_w4 = FFMIN(0, FFMIN(a, FFMIN(b, a + b)));
322+ const int max_h4 = FFMAX(0, FFMAX(c, FFMAX(d, c + d)));
323+ const int min_h4 = FFMIN(0, FFMIN(c, FFMIN(d, c + d)));
324+ const int bx_wx4 = ((max_w4 - min_w4) >> 11) + 9;
325+ const int bx_hx4 = ((max_h4 - min_h4) >> 11) + 9;
326+ return bx_wx4 * bx_hx4 > 225;
327+ } else {
328+ const int bx_wxh = (FFABS(a) >> 11) + 9;
329+ const int bx_hxh = (FFABS(d) >> 11) + 9;
330+ const int bx_wxv = (FFABS(b) >> 11) + 9;
331+ const int bx_hxv = (FFABS(c) >> 11) + 9;
332+ if (bx_wxh * bx_hxh <= 165 && bx_wxv * bx_hxv <= 165)
333+ return 0;
334+ }
335+ return 1;
336+}
337+
338+static void init_subblock_params(SubblockParams *sp, const MotionInfo* mi,
339+ const int cb_width, const int cb_height, const int lx)
340+{
341+ const int log2_cbw = av_log2(cb_width);
342+ const int log2_cbh = av_log2(cb_height);
343+ const Mv* cp_mv = mi->mv[lx];
344+ const int num_cp_mv = mi->motion_model_idc + 1;
345+ sp->d_hor_x = (cp_mv[1].x - cp_mv[0].x) << (MAX_CU_DEPTH - log2_cbw);
346+ sp->d_ver_x = (cp_mv[1].y - cp_mv[0].y) << (MAX_CU_DEPTH - log2_cbw);
347+ if (num_cp_mv == 3) {
348+ sp->d_hor_y = (cp_mv[2].x - cp_mv[0].x) << (MAX_CU_DEPTH - log2_cbh);
349+ sp->d_ver_y = (cp_mv[2].y - cp_mv[0].y) << (MAX_CU_DEPTH - log2_cbh);
350+ } else {
351+ sp->d_hor_y = -sp->d_ver_x;
352+ sp->d_ver_y = sp->d_hor_x;
353+ }
354+ sp->mv_scale_hor = (cp_mv[0].x) << MAX_CU_DEPTH;
355+ sp->mv_scale_ver = (cp_mv[0].y) << MAX_CU_DEPTH;
356+ sp->cb_width = cb_width;
357+ sp->cb_height = cb_height;
358+ sp->is_fallback = is_fallback_mode(sp, mi->pred_flag);
359+}
360+
361+static void derive_subblock_diff_mvs(const VVCLocalContext *lc, PredictionUnit* pu, const SubblockParams* sp, const int lx)
362+{
363+ pu->cb_prof_flag[lx] = derive_cb_prof_flag_lx(lc, pu, lx, sp->is_fallback);
364+ if (pu->cb_prof_flag[lx]) {
365+ const int dmv_limit = 1 << 5;
366+ const int pos_offset_x = 6 * (sp->d_hor_x + sp->d_hor_y);
367+ const int pos_offset_y = 6 * (sp->d_ver_x + sp->d_ver_y);
368+ for (int x = 0; x < AFFINE_MIN_BLOCK_SIZE; x++) {
369+ for (int y = 0; y < AFFINE_MIN_BLOCK_SIZE; y++) {
370+ Mv diff;
371+ diff.x = x * (sp->d_hor_x << 2) + y * (sp->d_hor_y << 2) - pos_offset_x;
372+ diff.y = x * (sp->d_ver_x << 2) + y * (sp->d_ver_y << 2) - pos_offset_y;
373+ ff_vvc_round_mv(&diff, 0, 8);
374+ pu->diff_mv_x[lx][AFFINE_MIN_BLOCK_SIZE * y + x] = av_clip(diff.x, -dmv_limit + 1, dmv_limit - 1);
375+ pu->diff_mv_y[lx][AFFINE_MIN_BLOCK_SIZE * y + x] = av_clip(diff.y, -dmv_limit + 1, dmv_limit - 1);
376+ }
377+ }
378+ }
379+}
380+
381+static void store_cp_mv(const VVCLocalContext *lc, const MotionInfo *mi, const int lx)
382+{
383+ VVCFrameContext *fc = lc->fc;
384+ const CodingUnit *cu = lc->cu;
385+ const int log2_min_cb_size = fc->ps.sps->min_cb_log2_size_y;
386+ const int min_cb_size = fc->ps.sps->min_cb_size_y;
387+ const int min_cb_width = fc->ps.pps->min_cb_width;
388+ const int num_cp_mv = mi->motion_model_idc + 1;
389+
390+ for (int dy = 0; dy < cu->cb_height; dy += min_cb_size) {
391+ for (int dx = 0; dx < cu->cb_width; dx += min_cb_size) {
392+ const int x_cb = (cu->x0 + dx) >> log2_min_cb_size;
393+ const int y_cb = (cu->y0 + dy) >> log2_min_cb_size;
394+ const int offset = (y_cb * min_cb_width + x_cb) * MAX_CONTROL_POINTS;
395+
396+ memcpy(&fc->tab.cp_mv[lx][offset], mi->mv[lx], sizeof(Mv) * num_cp_mv);
397+ SAMPLE_CTB(fc->tab.mmi, x_cb, y_cb) = mi->motion_model_idc;
398+ }
399+ }
400+}
401+
402+//8.5.5.9 Derivation process for motion vector arrays from affine control point motion vectors
403+void ff_vvc_store_sb_mvs(const VVCLocalContext *lc, PredictionUnit *pu)
404+{
405+ const CodingUnit *cu = lc->cu;
406+ const MotionInfo *mi = &pu->mi;
407+ const int sbw = cu->cb_width / mi->num_sb_x;
408+ const int sbh = cu->cb_height / mi->num_sb_y;
409+ SubblockParams params[2];
410+ MvField mvf;
411+
412+ mvf.pred_flag = mi->pred_flag;
413+ mvf.bcw_idx = mi->bcw_idx;
414+ mvf.hpel_if_idx = mi->hpel_if_idx;
415+ mvf.ciip_flag = 0;
416+ for (int i = 0; i < 2; i++) {
417+ const PredFlag mask = i + 1;
418+ if (mi->pred_flag & mask) {
419+ store_cp_mv(lc, mi, i);
420+ init_subblock_params(params + i, mi, cu->cb_width, cu->cb_height, i);
421+ derive_subblock_diff_mvs(lc, pu, params + i, i);
422+ mvf.ref_idx[i] = mi->ref_idx[i];
423+ }
424+ }
425+
426+ for (int sby = 0; sby < mi->num_sb_y; sby++) {
427+ for (int sbx = 0; sbx < mi->num_sb_x; sbx++) {
428+ const int x0 = cu->x0 + sbx * sbw;
429+ const int y0 = cu->y0 + sby * sbh;
430+ for (int i = 0; i < 2; i++) {
431+ const PredFlag mask = i + 1;
432+ if (mi->pred_flag & mask) {
433+ const SubblockParams* sp = params + i;
434+ const int x_pos_cb = sp->is_fallback ? (cu->cb_width >> 1) : (2 + (sbx << MIN_CU_LOG2));
435+ const int y_pos_cb = sp->is_fallback ? (cu->cb_height >> 1) : (2 + (sby << MIN_CU_LOG2));
436+ Mv *mv = mvf.mv + i;
437+
438+ mv->x = sp->mv_scale_hor + sp->d_hor_x * x_pos_cb + sp->d_hor_y * y_pos_cb;
439+ mv->y = sp->mv_scale_ver + sp->d_ver_x * x_pos_cb + sp->d_ver_y * y_pos_cb;
440+ ff_vvc_round_mv(mv, 0, MAX_CU_DEPTH);
441+ ff_vvc_clip_mv(mv);
442+ }
443+ }
444+ ff_vvc_set_mvf(lc, x0, y0, sbw, sbh, &mvf);
445+ }
446+ }
447+}
448+
449+void ff_vvc_store_gpm_mvf(const VVCLocalContext *lc, const PredictionUnit *pu)
450+{
451+ const CodingUnit *cu = lc->cu;
452+ const int angle_idx = ff_vvc_gpm_angle_idx[pu->gpm_partition_idx];
453+ const int distance_idx = ff_vvc_gpm_distance_idx[pu->gpm_partition_idx];
454+ const int displacement_x = ff_vvc_gpm_distance_lut[angle_idx];
455+ const int displacement_y = ff_vvc_gpm_distance_lut[(angle_idx + 8) % 32];
456+ const int is_flip = angle_idx >= 13 &&angle_idx <= 27;
457+ const int shift_hor = (angle_idx % 16 == 8 || (angle_idx % 16 && cu->cb_height >= cu->cb_width)) ? 0 : 1;
458+ const int sign = angle_idx < 16 ? 1 : -1;
459+ const int block_size = 4;
460+ int offset_x = (-cu->cb_width) >> 1;
461+ int offset_y = (-cu->cb_height) >> 1;
462+
463+ if (!shift_hor)
464+ offset_y += sign * ((distance_idx * cu->cb_height) >> 3);
465+ else
466+ offset_x += sign * ((distance_idx * cu->cb_width) >> 3);
467+
468+ for (int y = 0; y < cu->cb_height; y += block_size) {
469+ for (int x = 0; x < cu->cb_width; x += block_size) {
470+ const int motion_idx = (((x + offset_x) << 1) + 5) * displacement_x +
471+ (((y + offset_y) << 1) + 5) * displacement_y;
472+ const int s_type = FFABS(motion_idx) < 32 ? 2 : (motion_idx <= 0 ? (1 - is_flip) : is_flip);
473+ const int pred_flag = pu->gpm_mv[0].pred_flag | pu->gpm_mv[1].pred_flag;
474+ const int x0 = cu->x0 + x;
475+ const int y0 = cu->y0 + y;
476+
477+ if (!s_type)
478+ ff_vvc_set_mvf(lc, x0, y0, block_size, block_size, pu->gpm_mv + 0);
479+ else if (s_type == 1 || (s_type == 2 && pred_flag != PF_BI))
480+ ff_vvc_set_mvf(lc, x0, y0, block_size, block_size, pu->gpm_mv + 1);
481+ else {
482+ MvField mvf = pu->gpm_mv[0];
483+ const MvField *mv1 = &pu->gpm_mv[1];
484+ const int lx = mv1->pred_flag - PF_L0;
485+ mvf.pred_flag = PF_BI;
486+ mvf.ref_idx[lx] = mv1->ref_idx[lx];
487+ mvf.mv[lx] = mv1->mv[lx];
488+ ff_vvc_set_mvf(lc, x0, y0, block_size, block_size, &mvf);
489+ }
490+ }
491+ }
492+}
493+
494+void ff_vvc_store_mvf(const VVCLocalContext *lc, const MvField *mvf)
495+{
496+ const CodingUnit *cu = lc->cu;
497+ ff_vvc_set_mvf(lc, cu->x0, cu->y0, cu->cb_width, cu->cb_height, mvf);
498+}
499+
500+void ff_vvc_store_mv(const VVCLocalContext *lc, const MotionInfo *mi)
501+{
502+ const CodingUnit *cu = lc->cu;
503+ MvField mvf;
504+
505+ mvf.hpel_if_idx = mi->hpel_if_idx;
506+ mvf.bcw_idx = mi->bcw_idx;
507+ mvf.pred_flag = mi->pred_flag;
508+ mvf.ciip_flag = 0;
509+
510+ for (int i = 0; i < 2; i++) {
511+ const PredFlag mask = i + 1;
512+ if (mvf.pred_flag & mask) {
513+ mvf.mv[i] = mi->mv[i][0];
514+ mvf.ref_idx[i] = mi->ref_idx[i];
515+ }
516+ }
517+ ff_vvc_set_mvf(lc, cu->x0, cu->y0, cu->cb_width, cu->cb_height, &mvf);
518+}
519+
520+typedef enum NeighbourIdx {
521+ A0,
522+ A1,
523+ A2,
524+ B0,
525+ B1,
526+ B2,
527+ B3,
528+ NUM_NBS,
529+ NB_IDX_NONE = NUM_NBS,
530+} NeighbourIdx;
531+
532+typedef struct Neighbour {
533+ int x;
534+ int y;
535+
536+ int checked;
537+ int available;
538+} Neighbour;
539+
540+typedef struct NeighbourContext {
541+ Neighbour neighbours[NUM_NBS];
542+ const VVCLocalContext *lc;
543+} NeighbourContext;
544+
545+static int is_a0_available(const VVCLocalContext *lc, const CodingUnit *cu)
546+{
547+ const VVCFrameContext *fc = lc->fc;
548+ const VVCSPS *sps = fc->ps.sps;
549+ const int x0b = av_mod_uintp2(cu->x0, sps->ctb_log2_size_y);
550+ int cand_bottom_left;
551+
552+ if (!x0b && !lc->ctb_left_flag) {
553+ cand_bottom_left = 0;
554+ } else {
555+ const int log2_min_cb_size = sps->min_cb_log2_size_y;
556+ const int min_cb_width = fc->ps.pps->min_cb_width;
557+ const int x = (cu->x0 - 1) >> log2_min_cb_size;
558+ const int y = (cu->y0 + cu->cb_height) >> log2_min_cb_size;
559+ const int max_y = FFMIN(fc->ps.pps->height, ((cu->y0 >> sps->ctb_log2_size_y) + 1) << sps->ctb_log2_size_y);
560+ if (cu->y0 + cu->cb_height >= max_y)
561+ cand_bottom_left = 0;
562+ else
563+ cand_bottom_left = SAMPLE_CTB(fc->tab.cb_width[0], x, y) != 0;
564+ }
565+ return cand_bottom_left;
566+}
567+
568+static void init_neighbour_context(NeighbourContext *ctx, const VVCLocalContext *lc)
569+{
570+ const CodingUnit *cu = lc->cu;
571+ const NeighbourAvailable *na = &lc->na;
572+ const int x0 = cu->x0;
573+ const int y0 = cu->y0;
574+ const int cb_width = cu->cb_width;
575+ const int cb_height = cu->cb_height;
576+ const int a0_available = is_a0_available(lc, cu);
577+
578+ Neighbour neighbours[NUM_NBS] = {
579+ { x0 - 1, y0 + cb_height, !a0_available }, //A0
580+ { x0 - 1, y0 + cb_height - 1, !na->cand_left }, //A1
581+ { x0 - 1, y0, !na->cand_left }, //A2
582+ { x0 + cb_width, y0 - 1, !na->cand_up_right }, //B0
583+ { x0 + cb_width - 1, y0 - 1, !na->cand_up }, //B1
584+ { x0 - 1, y0 - 1, !na->cand_up_left }, //B2
585+ { x0, y0 - 1, !na->cand_up }, //B3
586+ };
587+
588+ memcpy(ctx->neighbours, neighbours, sizeof(neighbours));
589+ ctx->lc = lc;
590+}
591+
592+static int check_available(Neighbour *n, const VVCLocalContext *lc, const int is_mvp)
593+{
594+ const VVCFrameContext *fc = lc->fc;
595+ const VVCSPS *sps = fc->ps.sps;
596+ const CodingUnit *cu = lc->cu;
597+ const MvField *tab_mvf = fc->tab.mvf;
598+ const int min_pu_width = fc->ps.pps->min_pu_width;
599+
600+ if (!n->checked) {
601+ n->checked = 1;
602+ n->available = !sps->r->sps_entropy_coding_sync_enabled_flag || ((n->x >> sps->ctb_log2_size_y) <= (cu->x0 >> sps->ctb_log2_size_y));
603+ n->available &= TAB_MVF(n->x, n->y).pred_flag != PF_INTRA;
604+ if (!is_mvp)
605+ n->available &= !is_same_mer(fc, n->x, n->y, cu->x0, cu->y0);
606+ }
607+ return n->available;
608+}
609+
610+static const MvField *mv_merge_candidate(const VVCLocalContext *lc, const int x_cand, const int y_cand)
611+{
612+ const VVCFrameContext *fc = lc->fc;
613+ const int min_pu_width = fc->ps.pps->min_pu_width;
614+ const MvField* tab_mvf = fc->tab.mvf;
615+ const MvField *mvf = &TAB_MVF(x_cand, y_cand);
616+
617+ return mvf;
618+}
619+
620+static const MvField* mv_merge_from_nb(NeighbourContext *ctx, const NeighbourIdx nb)
621+{
622+ const VVCLocalContext *lc = ctx->lc;
623+ const int is_mvp = 0;
624+ Neighbour *n = &ctx->neighbours[nb];
625+
626+ if (check_available(n, lc, is_mvp))
627+ return mv_merge_candidate(lc, n->x, n->y);
628+ return 0;
629+}
630+#define MV_MERGE_FROM_NB(nb) mv_merge_from_nb(&nctx, nb)
631+
632+//8.5.2.3 Derivation process for spatial merging candidates
633+static int mv_merge_spatial_candidates(const VVCLocalContext *lc, const int merge_idx,
634+ const MvField **nb_list, MvField *cand_list, int *nb_merge_cand)
635+{
636+ const MvField *cand;
637+ int num_cands = 0;
638+ NeighbourContext nctx;
639+
640+ static NeighbourIdx nbs[][2] = {
641+ {B1, NB_IDX_NONE },
642+ {A1, B1 },
643+ {B0, B1 },
644+ {A0, A1 },
645+ };
646+
647+ init_neighbour_context(&nctx, lc);
648+ for (int i = 0; i < FF_ARRAY_ELEMS(nbs); i++) {
649+ NeighbourIdx nb = nbs[i][0];
650+ NeighbourIdx old = nbs[i][1];
651+ cand = nb_list[nb] = MV_MERGE_FROM_NB(nb);
652+ if (cand && !compare_mv_ref_idx(cand, nb_list[old])) {
653+ cand_list[num_cands] = *cand;
654+ if (merge_idx == num_cands)
655+ return 1;
656+ num_cands++;
657+ }
658+ }
659+ if (num_cands != 4) {
660+ cand = MV_MERGE_FROM_NB(B2);
661+ if (cand && !compare_mv_ref_idx(cand, nb_list[A1])
662+ && !compare_mv_ref_idx(cand, nb_list[B1])) {
663+ cand_list[num_cands] = *cand;
664+ if (merge_idx == num_cands)
665+ return 1;
666+ num_cands++;
667+ }
668+ }
669+ *nb_merge_cand = num_cands;
670+ return 0;
671+}
672+
673+static int mv_merge_temporal_candidate(const VVCLocalContext *lc, MvField *cand)
674+{
675+ const VVCFrameContext *fc = lc->fc;
676+ const CodingUnit *cu = lc->cu;
677+
678+ memset(cand, 0, sizeof(*cand));
679+ if (fc->ps.ph.r->ph_temporal_mvp_enabled_flag && (cu->cb_width * cu->cb_height > 32)) {
680+ int available_l0 = temporal_luma_motion_vector(lc, 0, cand->mv + 0, 0, 1, 0);
681+ int available_l1 = IS_B(lc->sc->sh.r) ?
682+ temporal_luma_motion_vector(lc, 0, cand->mv + 1, 1, 1, 0) : 0;
683+ cand->pred_flag = available_l0 + (available_l1 << 1);
684+ }
685+ return cand->pred_flag;
686+}
687+
688+//8.5.2.6 Derivation process for history-based merging candidates
689+static int mv_merge_history_candidates(const VVCLocalContext *lc, const int merge_idx,
690+ const MvField **nb_list, MvField *cand_list, int *num_cands)
691+{
692+ const VVCSPS *sps = lc->fc->ps.sps;
693+ const EntryPoint* ep = lc->ep;
694+ for (int i = 1; i <= ep->num_hmvp && (*num_cands < sps->max_num_merge_cand - 1); i++) {
695+ const MvField *h = &ep->hmvp[ep->num_hmvp - i];
696+ const int same_motion = i <= 2 && (compare_mv_ref_idx(h, nb_list[A1]) || compare_mv_ref_idx(h, nb_list[B1]));
697+ if (!same_motion) {
698+ cand_list[*num_cands] = *h;
699+ if (merge_idx == *num_cands)
700+ return 1;
701+ (*num_cands)++;
702+ }
703+ }
704+ return 0;
705+}
706+
707+//8.5.2.4 Derivation process for pairwise average merging candidate
708+static int mv_merge_pairwise_candidate(MvField *cand_list, const int num_cands, const int is_b)
709+{
710+ if (num_cands > 1) {
711+ const int num_ref_rists = is_b ? 2 : 1;
712+ const MvField* p0 = cand_list + 0;
713+ const MvField* p1 = cand_list + 1;
714+ MvField* cand = cand_list + num_cands;
715+
716+ cand->pred_flag = 0;
717+ for (int i = 0; i < num_ref_rists; i++) {
718+ PredFlag mask = i + 1;
719+ if (p0->pred_flag & mask) {
720+ cand->pred_flag |= mask;
721+ cand->ref_idx[i] = p0->ref_idx[i];
722+ if (p1->pred_flag & mask) {
723+ Mv *mv = cand->mv + i;
724+ mv->x = p0->mv[i].x + p1->mv[i].x;
725+ mv->y = p0->mv[i].y + p1->mv[i].y;
726+ ff_vvc_round_mv(mv, 0, 1);
727+ } else {
728+ cand->mv[i] = p0->mv[i];
729+ }
730+ } else if (p1->pred_flag & mask) {
731+ cand->pred_flag |= mask;
732+ cand->mv[i] = p1->mv[i];
733+ cand->ref_idx[i] = p1->ref_idx[i];
734+ }
735+ }
736+ if (cand->pred_flag) {
737+ cand->hpel_if_idx = p0->hpel_if_idx == p1->hpel_if_idx ? p0->hpel_if_idx : 0;
738+ cand->bcw_idx = 0;
739+ cand->ciip_flag = 0;
740+ return 1;
741+ }
742+ }
743+ return 0;
744+}
745+
746+//8.5.2.5 Derivation process for zero motion vector merging candidates
747+static void mv_merge_zero_motion_candidate(const VVCLocalContext *lc, const int merge_idx,
748+ MvField *cand_list, int num_cands)
749+{
750+ const VVCSPS *sps = lc->fc->ps.sps;
751+ const H266RawSliceHeader *rsh = lc->sc->sh.r;
752+ const int num_ref_idx = IS_P(rsh) ?
753+ rsh->num_ref_idx_active[L0] : FFMIN(rsh->num_ref_idx_active[L0], rsh->num_ref_idx_active[L1]);
754+ int zero_idx = 0;
755+
756+ while (num_cands < sps->max_num_merge_cand) {
757+ MvField *cand = cand_list + num_cands;
758+
759+ cand->pred_flag = PF_L0 + (IS_B(rsh) << 1);
760+ AV_ZERO64(cand->mv + 0);
761+ AV_ZERO64(cand->mv + 1);
762+ cand->ref_idx[0] = zero_idx < num_ref_idx ? zero_idx : 0;
763+ cand->ref_idx[1] = zero_idx < num_ref_idx ? zero_idx : 0;
764+ cand->bcw_idx = 0;
765+ cand->hpel_if_idx = 0;
766+ if (merge_idx == num_cands)
767+ return;
768+ num_cands++;
769+ zero_idx++;
770+ }
771+}
772+
773+static void mv_merge_mode(const VVCLocalContext *lc, const int merge_idx, MvField *cand_list)
774+{
775+ int num_cands = 0;
776+ const MvField *nb_list[NUM_NBS + 1] = { NULL };
777+
778+ if (mv_merge_spatial_candidates(lc, merge_idx, nb_list, cand_list, &num_cands))
779+ return;
780+
781+ if (mv_merge_temporal_candidate(lc, &cand_list[num_cands])) {
782+ if (merge_idx == num_cands)
783+ return;
784+ num_cands++;
785+ }
786+
787+ if (mv_merge_history_candidates(lc, merge_idx, nb_list, cand_list, &num_cands))
788+ return;
789+
790+ if (mv_merge_pairwise_candidate(cand_list, num_cands, IS_B(lc->sc->sh.r))) {
791+ if (merge_idx == num_cands)
792+ return;
793+ num_cands++;
794+ }
795+
796+ mv_merge_zero_motion_candidate(lc, merge_idx, cand_list, num_cands);
797+}
798+
799+//8.5.2.2 Derivation process for luma motion vectors for merge mode
800+void ff_vvc_luma_mv_merge_mode(VVCLocalContext *lc, const int merge_idx, const int ciip_flag, MvField *mv)
801+{
802+ const CodingUnit *cu = lc->cu;
803+ MvField cand_list[MRG_MAX_NUM_CANDS];
804+
805+ ff_vvc_set_neighbour_available(lc, cu->x0, cu->y0, cu->cb_width, cu->cb_height);
806+ mv_merge_mode(lc, merge_idx, cand_list);
807+ *mv = cand_list[merge_idx];
808+ //ciip flag in not inhritable
809+ mv->ciip_flag = ciip_flag;
810+}
811+
812+//8.5.4.2 Derivation process for luma motion vectors for geometric partitioning merge mode
813+void ff_vvc_luma_mv_merge_gpm(VVCLocalContext *lc, const int merge_gpm_idx[2], MvField *mv)
814+{
815+ const CodingUnit *cu = lc->cu;
816+ MvField cand_list[MRG_MAX_NUM_CANDS];
817+
818+ const int idx[] = { merge_gpm_idx[0], merge_gpm_idx[1] + (merge_gpm_idx[1] >= merge_gpm_idx[0]) };
819+
820+ ff_vvc_set_neighbour_available(lc, cu->x0, cu->y0, cu->cb_width, cu->cb_height);
821+ mv_merge_mode(lc, FFMAX(idx[0], idx[1]), cand_list);
822+ memset(mv, 0, 2 * sizeof(*mv));
823+ for (int i = 0; i < 2; i++) {
824+ int lx = idx[i] & 1;
825+ int mask = lx + PF_L0;
826+ MvField *cand = cand_list + idx[i];
827+ if (!(cand->pred_flag & mask)) {
828+ lx = !lx;
829+ mask = lx + PF_L0;
830+ }
831+ mv[i].pred_flag = mask;
832+ mv[i].ref_idx[lx] = cand->ref_idx[lx];
833+ mv[i].mv[lx] = cand->mv[lx];
834+ }
835+
836+}
837+
838+//8.5.5.5 Derivation process for luma affine control point motion vectors from a neighbouring block
839+static void affine_cps_from_nb(const VVCLocalContext *lc,
840+ const int x_nb, int y_nb, const int nbw, const int nbh, const int lx,
841+ Mv *cps, int num_cps)
842+{
843+ const VVCFrameContext *fc = lc->fc;
844+ const CodingUnit *cu = lc->cu;
845+ const int x0 = cu->x0;
846+ const int y0 = cu->y0;
847+ const int cb_width = cu->cb_width;
848+ const int cb_height = cu->cb_height;
849+ const MvField* tab_mvf = fc->tab.mvf;
850+ const int min_cb_log2_size = fc->ps.sps->min_cb_log2_size_y;
851+ const int min_cb_width = fc->ps.pps->min_cb_width;
852+
853+ const int log2_nbw = ff_log2(nbw);
854+ const int log2_nbh = ff_log2(nbh);
855+ const int is_ctb_boundary = !((y_nb + nbh) % fc->ps.sps->ctb_size_y) && (y_nb + nbh == y0);
856+ const Mv *l, *r;
857+ int mv_scale_hor, mv_scale_ver, d_hor_x, d_ver_x, d_hor_y, d_ver_y, motion_model_idc_nb;
858+ if (is_ctb_boundary) {
859+ const int min_pu_width = fc->ps.pps->min_pu_width;
860+ l = &TAB_MVF(x_nb, y_nb + nbh - 1).mv[lx];
861+ r = &TAB_MVF(x_nb + nbw - 1, y_nb + nbh - 1).mv[lx];
862+ } else {
863+ const int x = x_nb >> min_cb_log2_size;
864+ const int y = y_nb >> min_cb_log2_size;
865+ motion_model_idc_nb = SAMPLE_CTB(fc->tab.mmi, x, y);
866+
867+ l = &TAB_CP_MV(lx, x_nb, y_nb);
868+ r = &TAB_CP_MV(lx, x_nb + nbw - 1, y_nb) + 1;
869+ }
870+ mv_scale_hor = l->x << 7;
871+ mv_scale_ver = l->y << 7;
872+ d_hor_x = (r->x - l->x) << (7 - log2_nbw);
873+ d_ver_x = (r->y - l->y) << (7 - log2_nbw);
874+ if (!is_ctb_boundary && motion_model_idc_nb == MOTION_6_PARAMS_AFFINE) {
875+ const Mv* lb = &TAB_CP_MV(lx, x_nb, y_nb + nbh - 1) + 2;
876+ d_hor_y = (lb->x - l->x) << (7 - log2_nbh);
877+ d_ver_y = (lb->y - l->y) << (7 - log2_nbh);
878+ } else {
879+ d_hor_y = -d_ver_x;
880+ d_ver_y = d_hor_x;
881+ }
882+
883+ if (is_ctb_boundary) {
884+ y_nb = y0;
885+ }
886+ cps[0].x = mv_scale_hor + d_hor_x * (x0 - x_nb) + d_hor_y * (y0 - y_nb);
887+ cps[0].y = mv_scale_ver + d_ver_x * (x0 - x_nb) + d_ver_y * (y0 - y_nb);
888+ cps[1].x = mv_scale_hor + d_hor_x * (x0 + cb_width - x_nb) + d_hor_y * (y0 - y_nb);
889+ cps[1].y = mv_scale_ver + d_ver_x * (x0 + cb_width - x_nb) + d_ver_y * (y0 - y_nb);
890+ if (num_cps == 3) {
891+ cps[2].x = mv_scale_hor + d_hor_x * (x0 - x_nb) + d_hor_y * (y0 + cb_height - y_nb);
892+ cps[2].y = mv_scale_ver + d_ver_x * (x0 - x_nb) + d_ver_y * (y0 + cb_height - y_nb);
893+ }
894+ for (int i = 0; i < num_cps; i++) {
895+ ff_vvc_round_mv(cps + i, 0, 7);
896+ ff_vvc_clip_mv(cps + i);
897+ }
898+}
899+
900+//derive affine neighbour's postion, width and height,
901+static int affine_neighbour_cb(const VVCFrameContext *fc, const int x_nb, const int y_nb, int *x_cb, int *y_cb, int *cbw, int *cbh)
902+{
903+ const int log2_min_cb_size = fc->ps.sps->min_cb_log2_size_y;
904+ const int min_cb_width = fc->ps.pps->min_cb_width;
905+ const int x = x_nb >> log2_min_cb_size;
906+ const int y = y_nb >> log2_min_cb_size;
907+ const int motion_model_idc = SAMPLE_CTB(fc->tab.mmi, x, y);
908+ if (motion_model_idc) {
909+ *x_cb = SAMPLE_CTB(fc->tab.cb_pos_x[0], x, y);
910+ *y_cb = SAMPLE_CTB(fc->tab.cb_pos_y[0], x, y);
911+ *cbw = SAMPLE_CTB(fc->tab.cb_width[0], x, y);
912+ *cbh = SAMPLE_CTB(fc->tab.cb_height[0], x, y);
913+ }
914+ return motion_model_idc;
915+}
916+
917+//part of 8.5.5.2 Derivation process for motion vectors and reference indices in subblock merge mode
918+static int affine_merge_candidate(const VVCLocalContext *lc, const int x_cand, const int y_cand, MotionInfo* mi)
919+{
920+ const VVCFrameContext *fc = lc->fc;
921+ int x, y, w, h, motion_model_idc;
922+
923+ motion_model_idc = affine_neighbour_cb(fc, x_cand, y_cand, &x, &y, &w, &h);
924+ if (motion_model_idc) {
925+ const int min_pu_width = fc->ps.pps->min_pu_width;
926+ const MvField* tab_mvf = fc->tab.mvf;
927+ const MvField *mvf = &TAB_MVF(x, y);
928+
929+ mi->bcw_idx = mvf->bcw_idx;
930+ mi->pred_flag = mvf->pred_flag;
931+ for (int i = 0; i < 2; i++) {
932+ PredFlag mask = i + 1;
933+ if (mi->pred_flag & mask) {
934+ affine_cps_from_nb(lc, x, y, w, h, i, &mi->mv[i][0], motion_model_idc + 1);
935+ }
936+ mi->ref_idx[i] = mvf->ref_idx[i];
937+ }
938+ mi->motion_model_idc = motion_model_idc;
939+ }
940+ return motion_model_idc;
941+}
942+
943+static int affine_merge_from_nbs(NeighbourContext *ctx, const NeighbourIdx *nbs, const int num_nbs, MotionInfo* cand)
944+{
945+ const VVCLocalContext *lc = ctx->lc;
946+ const int is_mvp = 0;
947+ for (int i = 0; i < num_nbs; i++) {
948+ Neighbour *n = &ctx->neighbours[nbs[i]];
949+ if (check_available(n, lc, is_mvp) && affine_merge_candidate(lc, n->x, n->y, cand))
950+ return 1;
951+ }
952+ return 0;
953+}
954+#define AFFINE_MERGE_FROM_NBS(nbs) affine_merge_from_nbs(&nctx, nbs, FF_ARRAY_ELEMS(nbs), mi)
955+
956+
957+static const MvField* derive_corner_mvf(NeighbourContext *ctx, const NeighbourIdx *neighbour, const int num_neighbour)
958+{
959+ const VVCFrameContext *fc = ctx->lc->fc;
960+ const MvField *tab_mvf = fc->tab.mvf;
961+ const int min_pu_width = fc->ps.pps->min_pu_width;
962+ for (int i = 0; i < num_neighbour; i++) {
963+ Neighbour *n = &ctx->neighbours[neighbour[i]];
964+ if (check_available(n, ctx->lc, 0)) {
965+ return &TAB_MVF(n->x, n->y);
966+ }
967+ }
968+ return NULL;
969+}
970+
971+#define DERIVE_CORNER_MV(nbs) derive_corner_mvf(nctx, nbs, FF_ARRAY_ELEMS(nbs))
972+
973+// check if the mv's and refidx are the same between A and B
974+static av_always_inline int compare_pf_ref_idx(const MvField *A, const struct MvField *B, const struct MvField *C, const int lx)
975+{
976+
977+ const PredFlag mask = (lx + 1) & A->pred_flag;
978+ if (!(B->pred_flag & mask))
979+ return 0;
980+ if (A->ref_idx[lx] != B->ref_idx[lx])
981+ return 0;
982+ if (C) {
983+ if (!(C->pred_flag & mask))
984+ return 0;
985+ if (A->ref_idx[lx] != C->ref_idx[lx])
986+ return 0;
987+ }
988+ return 1;
989+}
990+
991+static av_always_inline void sb_clip_location(const VVCFrameContext *fc,
992+ const int x_ctb, const int y_ctb, const Mv* temp_mv, int *x, int *y)
993+{
994+ const VVCPPS *pps = fc->ps.pps;
995+ const int ctb_log2_size = fc->ps.sps->ctb_log2_size_y;
996+ *y = av_clip(*y + temp_mv->y, y_ctb, FFMIN(pps->height - 1, y_ctb + (1 << ctb_log2_size) - 1)) & ~7;
997+ *x = av_clip(*x + temp_mv->x, x_ctb, FFMIN(pps->width - 1, x_ctb + (1 << ctb_log2_size) + 3)) & ~7;
998+}
999+
1000+static void sb_temproal_luma_motion(const VVCLocalContext *lc,
1001+ const int x_ctb, const int y_ctb, const Mv *temp_mv,
1002+ int x, int y, uint8_t *pred_flag, Mv *mv)
1003+{
1004+ MvField temp_col;
1005+ Mv* mvLXCol;
1006+ const int refIdxLx = 0;
1007+ const VVCFrameContext *fc = lc->fc;
1008+ const VVCSH *sh = &lc->sc->sh;
1009+ const int min_pu_width = fc->ps.pps->min_pu_width;
1010+ VVCFrame *ref = fc->ref->collocated_ref;
1011+ MvField *tab_mvf = ref->tab_dmvr_mvf;
1012+ int colPic = ref->poc;
1013+ int X = 0;
1014+
1015+ sb_clip_location(fc, x_ctb, y_ctb, temp_mv, &x, &y);
1016+
1017+ temp_col = TAB_MVF(x, y);
1018+ mvLXCol = mv + 0;
1019+ *pred_flag = DERIVE_TEMPORAL_COLOCATED_MVS(1);
1020+ if (IS_B(sh->r)) {
1021+ X = 1;
1022+ mvLXCol = mv + 1;
1023+ *pred_flag |= (DERIVE_TEMPORAL_COLOCATED_MVS(1)) << 1;
1024+ }
1025+}
1026+
1027+//8.5.5.4 Derivation process for subblock-based temporal merging base motion data
1028+static int sb_temporal_luma_motion_data(const VVCLocalContext *lc, const MvField *a1,
1029+ const int x_ctb, const int y_ctb, MvField *ctr_mvf, Mv *temp_mv)
1030+{
1031+ const VVCFrameContext *fc = lc->fc;
1032+ const RefPicList *rpl = lc->sc->rpl;
1033+ const CodingUnit *cu = lc->cu;
1034+ const int x = cu->x0 + cu->cb_width / 2;
1035+ const int y = cu->y0 + cu->cb_height / 2;
1036+ const VVCFrame *ref = fc->ref->collocated_ref;
1037+
1038+ int colPic;
1039+
1040+ memset(temp_mv, 0, sizeof(*temp_mv));
1041+
1042+ if (!ref) {
1043+ memset(ctr_mvf, 0, sizeof(*ctr_mvf));
1044+ return 0;
1045+ }
1046+
1047+ colPic = ref->poc;
1048+
1049+ AV_ZERO64(temp_mv);
1050+ if (a1) {
1051+ if ((a1->pred_flag & PF_L0) && colPic == rpl[0].list[a1->ref_idx[0]])
1052+ *temp_mv = a1->mv[0];
1053+ else if ((a1->pred_flag & PF_L1) && colPic == rpl[1].list[a1->ref_idx[1]])
1054+ *temp_mv = a1->mv[1];
1055+ ff_vvc_round_mv(temp_mv, 0, 4);
1056+ }
1057+ sb_temproal_luma_motion(lc, x_ctb, y_ctb, temp_mv, x, y, &ctr_mvf->pred_flag , ctr_mvf->mv);
1058+
1059+ return ctr_mvf->pred_flag;
1060+}
1061+
1062+
1063+//8.5.5.3 Derivation process for subblock-based temporal merging candidates
1064+static int sb_temporal_merge_candidate(const VVCLocalContext* lc, NeighbourContext *nctx, PredictionUnit *pu)
1065+{
1066+ const VVCFrameContext *fc = lc->fc;
1067+ const CodingUnit *cu = lc->cu;
1068+ const VVCSPS *sps = fc->ps.sps;
1069+ const VVCPH *ph = &fc->ps.ph;
1070+ MotionInfo *mi = &pu->mi;
1071+ const int ctb_log2_size = sps->ctb_log2_size_y;
1072+ const int x0 = cu->x0;
1073+ const int y0 = cu->y0;
1074+ const NeighbourIdx n = A1;
1075+ const MvField *a1;
1076+ MvField ctr_mvf;
1077+ Mv temp_mv;
1078+ const int x_ctb = (x0 >> ctb_log2_size) << ctb_log2_size;
1079+ const int y_ctb = (y0 >> ctb_log2_size) << ctb_log2_size;
1080+
1081+
1082+ if (!ph->r->ph_temporal_mvp_enabled_flag ||
1083+ !sps->r->sps_sbtmvp_enabled_flag ||
1084+ (cu->cb_width < 8 && cu->cb_height < 8))
1085+ return 0;
1086+
1087+ mi->num_sb_x = cu->cb_width >> 3;
1088+ mi->num_sb_y = cu->cb_height >> 3;
1089+
1090+ a1 = derive_corner_mvf(nctx, &n, 1);
1091+ if (sb_temporal_luma_motion_data(lc, a1, x_ctb, y_ctb, &ctr_mvf, &temp_mv)) {
1092+ const int sbw = cu->cb_width / mi->num_sb_x;
1093+ const int sbh = cu->cb_height / mi->num_sb_y;
1094+ MvField mvf = {0};
1095+ for (int sby = 0; sby < mi->num_sb_y; sby++) {
1096+ for (int sbx = 0; sbx < mi->num_sb_x; sbx++) {
1097+ int x = x0 + sbx * sbw;
1098+ int y = y0 + sby * sbh;
1099+ sb_temproal_luma_motion(lc, x_ctb, y_ctb, &temp_mv, x + sbw / 2, y + sbh / 2, &mvf.pred_flag, mvf.mv);
1100+ if (!mvf.pred_flag) {
1101+ mvf.pred_flag = ctr_mvf.pred_flag;
1102+ memcpy(mvf.mv, ctr_mvf.mv, sizeof(mvf.mv));
1103+ }
1104+ ff_vvc_set_mvf(lc, x, y, sbw, sbh, &mvf);
1105+ }
1106+ }
1107+ return 1;
1108+ }
1109+ return 0;
1110+}
1111+
1112+static int affine_merge_const1(const MvField *c0, const MvField *c1, const MvField *c2, MotionInfo *mi)
1113+{
1114+ if (c0 && c1 && c2) {
1115+ mi->pred_flag = 0;
1116+ for (int i = 0; i < 2; i++) {
1117+ PredFlag mask = i + 1;
1118+ if (compare_pf_ref_idx(c0, c1, c2, i)) {
1119+ mi->pred_flag |= mask;
1120+ mi->ref_idx[i] = c0->ref_idx[i];
1121+ mi->mv[i][0] = c0->mv[i];
1122+ mi->mv[i][1] = c1->mv[i];
1123+ mi->mv[i][2] = c2->mv[i];
1124+ }
1125+ }
1126+ if (mi->pred_flag) {
1127+ if (mi->pred_flag == PF_BI)
1128+ mi->bcw_idx = c0->bcw_idx;
1129+ mi->motion_model_idc = MOTION_6_PARAMS_AFFINE;
1130+ return 1;
1131+ }
1132+ }
1133+ return 0;
1134+}
1135+
1136+static int affine_merge_const2(const MvField *c0, const MvField *c1, const MvField *c3, MotionInfo *mi)
1137+{
1138+ if (c0 && c1 && c3) {
1139+ mi->pred_flag = 0;
1140+ for (int i = 0; i < 2; i++) {
1141+ PredFlag mask = i + 1;
1142+ if (compare_pf_ref_idx(c0, c1, c3, i)) {
1143+ mi->pred_flag |= mask;
1144+ mi->ref_idx[i] = c0->ref_idx[i];
1145+ mi->mv[i][0] = c0->mv[i];
1146+ mi->mv[i][1] = c1->mv[i];
1147+ mi->mv[i][2].x = c3->mv[i].x + c0->mv[i].x - c1->mv[i].x;
1148+ mi->mv[i][2].y = c3->mv[i].y + c0->mv[i].y - c1->mv[i].y;
1149+ ff_vvc_clip_mv(&mi->mv[i][2]);
1150+ }
1151+ }
1152+ if (mi->pred_flag) {
1153+ mi->bcw_idx = mi->pred_flag == PF_BI ? c0->bcw_idx : 0;
1154+ mi->motion_model_idc = MOTION_6_PARAMS_AFFINE;
1155+ return 1;
1156+ }
1157+ }
1158+ return 0;
1159+}
1160+
1161+static int affine_merge_const3(const MvField *c0, const MvField *c2, const MvField *c3, MotionInfo *mi)
1162+{
1163+ if (c0 && c2 && c3) {
1164+ mi->pred_flag = 0;
1165+ for (int i = 0; i < 2; i++) {
1166+ PredFlag mask = i + 1;
1167+ if (compare_pf_ref_idx(c0, c2, c3, i)) {
1168+ mi->pred_flag |= mask;
1169+ mi->ref_idx[i] = c0->ref_idx[i];
1170+ mi->mv[i][0] = c0->mv[i];
1171+ mi->mv[i][1].x = c3->mv[i].x + c0->mv[i].x - c2->mv[i].x;
1172+ mi->mv[i][1].y = c3->mv[i].y + c0->mv[i].y - c2->mv[i].y;
1173+ ff_vvc_clip_mv(&mi->mv[i][1]);
1174+ mi->mv[i][2] = c2->mv[i];
1175+ }
1176+ }
1177+ if (mi->pred_flag) {
1178+ mi->bcw_idx = mi->pred_flag == PF_BI ? c0->bcw_idx : 0;
1179+ mi->motion_model_idc = MOTION_6_PARAMS_AFFINE;
1180+ return 1;
1181+ }
1182+ }
1183+ return 0;
1184+}
1185+
1186+static int affine_merge_const4(const MvField *c1, const MvField *c2, const MvField *c3, MotionInfo *mi)
1187+{
1188+ if (c1 && c2 && c3) {
1189+ mi->pred_flag = 0;
1190+ for (int i = 0; i < 2; i++) {
1191+ PredFlag mask = i + 1;
1192+ if (compare_pf_ref_idx(c1, c2, c3, i)) {
1193+ mi->pred_flag |= mask;
1194+ mi->ref_idx[i] = c1->ref_idx[i];
1195+ mi->mv[i][0].x = c1->mv[i].x + c2->mv[i].x - c3->mv[i].x;
1196+ mi->mv[i][0].y = c1->mv[i].y + c2->mv[i].y - c3->mv[i].y;
1197+ ff_vvc_clip_mv(&mi->mv[i][0]);
1198+ mi->mv[i][1] = c1->mv[i];
1199+ mi->mv[i][2] = c2->mv[i];
1200+ }
1201+ }
1202+ if (mi->pred_flag) {
1203+ mi->bcw_idx = mi->pred_flag == PF_BI ? c1->bcw_idx : 0;
1204+ mi->motion_model_idc = MOTION_6_PARAMS_AFFINE;
1205+ return 1;
1206+ }
1207+ }
1208+ return 0;
1209+}
1210+
1211+static int affine_merge_const5(const MvField *c0, const MvField *c1, MotionInfo *mi)
1212+{
1213+ if (c0 && c1) {
1214+ mi->pred_flag = 0;
1215+ for (int i = 0; i < 2; i++) {
1216+ PredFlag mask = i + 1;
1217+ if (compare_pf_ref_idx(c0, c1, NULL, i)) {
1218+ mi->pred_flag |= mask;
1219+ mi->ref_idx[i] = c0->ref_idx[i];
1220+ mi->mv[i][0] = c0->mv[i];
1221+ mi->mv[i][1] = c1->mv[i];
1222+ }
1223+ }
1224+ if (mi->pred_flag) {
1225+ if (mi->pred_flag == PF_BI)
1226+ mi->bcw_idx = c0->bcw_idx;
1227+ mi->motion_model_idc = MOTION_4_PARAMS_AFFINE;
1228+ return 1;
1229+ }
1230+ }
1231+ return 0;
1232+}
1233+
1234+static int affine_merge_const6(const MvField* c0, const MvField* c2, const int cb_width, const int cb_height, MotionInfo *mi)
1235+{
1236+ if (c0 && c2) {
1237+ const int shift = 7 + av_log2(cb_width) - av_log2(cb_height);
1238+ mi->pred_flag = 0;
1239+ for (int i = 0; i < 2; i++) {
1240+ PredFlag mask = i + 1;
1241+ if (compare_pf_ref_idx(c0, c2, NULL, i)) {
1242+ mi->pred_flag |= mask;
1243+ mi->ref_idx[i] = c0->ref_idx[i];
1244+ mi->mv[i][0] = c0->mv[i];
1245+ mi->mv[i][1].x = (c0->mv[i].x << 7) + ((c2->mv[i].y - c0->mv[i].y) << shift);
1246+ mi->mv[i][1].y = (c0->mv[i].y << 7) - ((c2->mv[i].x - c0->mv[i].x) << shift);
1247+ ff_vvc_round_mv(&mi->mv[i][1], 0, 7);
1248+ ff_vvc_clip_mv(&mi->mv[i][1]);
1249+ }
1250+ }
1251+ if (mi->pred_flag) {
1252+ if (mi->pred_flag == PF_BI)
1253+ mi->bcw_idx = c0->bcw_idx;
1254+ mi->motion_model_idc = MOTION_4_PARAMS_AFFINE;
1255+ return 1;
1256+ }
1257+ }
1258+ return 0;
1259+}
1260+
1261+static void affine_merge_zero_motion(const VVCLocalContext *lc, MotionInfo *mi)
1262+{
1263+ const CodingUnit *cu = lc->cu;
1264+
1265+ memset(mi, 0, sizeof(*mi));
1266+ mi->pred_flag = PF_L0 + (IS_B(lc->sc->sh.r) << 1);
1267+ mi->motion_model_idc = MOTION_4_PARAMS_AFFINE;
1268+ mi->num_sb_x = cu->cb_width >> MIN_PU_LOG2;
1269+ mi->num_sb_y = cu->cb_height >> MIN_PU_LOG2;
1270+}
1271+
1272+//8.5.5.6 Derivation process for constructed affine control point motion vector merging candidates
1273+static int affine_merge_const_candidates(const VVCLocalContext *lc, MotionInfo *mi,
1274+ NeighbourContext *nctx, const int merge_subblock_idx, int num_cands)
1275+{
1276+ const VVCFrameContext *fc = lc->fc;
1277+ const CodingUnit *cu = lc->cu;
1278+ const NeighbourIdx tl[] = { B2, B3, A2 };
1279+ const NeighbourIdx tr[] = { B1, B0};
1280+ const NeighbourIdx bl[] = { A1, A0};
1281+ const MvField *c0, *c1, *c2;
1282+
1283+ c0 = DERIVE_CORNER_MV(tl);
1284+ c1 = DERIVE_CORNER_MV(tr);
1285+ c2 = DERIVE_CORNER_MV(bl);
1286+
1287+ if (fc->ps.sps->r->sps_6param_affine_enabled_flag) {
1288+ MvField corner3, *c3 = NULL;
1289+ //Const1
1290+ if (affine_merge_const1(c0, c1, c2, mi)) {
1291+ if (merge_subblock_idx == num_cands)
1292+ return 1;
1293+ num_cands++;
1294+ }
1295+
1296+ memset(&corner3, 0, sizeof(corner3));
1297+ if (fc->ps.ph.r->ph_temporal_mvp_enabled_flag){
1298+ const int available_l0 = temporal_luma_motion_vector(lc, 0, corner3.mv + 0, 0, 0, 0);
1299+ const int available_l1 = (lc->sc->sh.r->sh_slice_type == VVC_SLICE_TYPE_B) ?
1300+ temporal_luma_motion_vector(lc, 0, corner3.mv + 1, 1, 0, 0) : 0;
1301+
1302+ corner3.pred_flag = available_l0 + (available_l1 << 1);
1303+ if (corner3.pred_flag)
1304+ c3 = &corner3;
1305+ }
1306+
1307+ //Const2
1308+ if (affine_merge_const2(c0, c1, c3, mi)) {
1309+ if (merge_subblock_idx == num_cands)
1310+ return 1;
1311+ num_cands++;
1312+ }
1313+
1314+ //Const3
1315+ if (affine_merge_const3(c0, c2, c3, mi)) {
1316+ if (merge_subblock_idx == num_cands)
1317+ return 1;
1318+ num_cands++;
1319+ }
1320+
1321+ //Const4
1322+ if (affine_merge_const4(c1, c2, c3, mi)) {
1323+ if (merge_subblock_idx == num_cands)
1324+ return 1;
1325+ num_cands++;
1326+ }
1327+ }
1328+
1329+ //Const5
1330+ if (affine_merge_const5(c0, c1, mi)) {
1331+ if (merge_subblock_idx == num_cands)
1332+ return 1;
1333+ num_cands++;
1334+ }
1335+
1336+ if (affine_merge_const6(c0, c2, cu->cb_width, cu->cb_height, mi)) {
1337+ if (merge_subblock_idx == num_cands)
1338+ return 1;
1339+ }
1340+ return 0;
1341+}
1342+
1343+//8.5.5.2 Derivation process for motion vectors and reference indices in subblock merge mode
1344+//return 1 if candidate is SbCol
1345+static int sb_mv_merge_mode(const VVCLocalContext *lc, const int merge_subblock_idx, PredictionUnit *pu)
1346+{
1347+ const VVCSPS *sps = lc->fc->ps.sps;
1348+ const CodingUnit *cu = lc->cu;
1349+ MotionInfo *mi = &pu->mi;
1350+ int num_cands = 0;
1351+ NeighbourContext nctx;
1352+
1353+ init_neighbour_context(&nctx, lc);
1354+
1355+ //SbCol
1356+ if (sb_temporal_merge_candidate(lc, &nctx, pu)) {
1357+ if (merge_subblock_idx == num_cands)
1358+ return 1;
1359+ num_cands++;
1360+ }
1361+
1362+ pu->inter_affine_flag = 1;
1363+ mi->num_sb_x = cu->cb_width >> MIN_PU_LOG2;
1364+ mi->num_sb_y = cu->cb_height >> MIN_PU_LOG2;
1365+
1366+ if (sps->r->sps_affine_enabled_flag) {
1367+ const NeighbourIdx ak[] = { A0, A1 };
1368+ const NeighbourIdx bk[] = { B0, B1, B2 };
1369+ //A
1370+ if (AFFINE_MERGE_FROM_NBS(ak)) {
1371+ if (merge_subblock_idx == num_cands)
1372+ return 0;
1373+ num_cands++;
1374+ }
1375+
1376+ //B
1377+ if (AFFINE_MERGE_FROM_NBS(bk)) {
1378+ if (merge_subblock_idx == num_cands)
1379+ return 0;
1380+ num_cands++;
1381+ }
1382+
1383+ //Const1 to Const6
1384+ if (affine_merge_const_candidates(lc, mi, &nctx, merge_subblock_idx, num_cands))
1385+ return 0;
1386+ }
1387+ //Zero
1388+ affine_merge_zero_motion(lc, mi);
1389+ return 0;
1390+}
1391+
1392+void ff_vvc_sb_mv_merge_mode(VVCLocalContext *lc, const int merge_subblock_idx, PredictionUnit *pu)
1393+{
1394+ const CodingUnit *cu = lc->cu;
1395+ ff_vvc_set_neighbour_available(lc, cu->x0, cu->y0, cu->cb_width, cu->cb_height);
1396+ if (!sb_mv_merge_mode(lc, merge_subblock_idx, pu)) {
1397+ ff_vvc_store_sb_mvs(lc, pu);
1398+ }
1399+}
1400+
1401+static int mvp_candidate(const VVCLocalContext *lc, const int x_cand, const int y_cand,
1402+ const int lx, const int8_t *ref_idx, Mv *mv)
1403+{
1404+ const VVCFrameContext *fc = lc->fc;
1405+ const RefPicList *rpl = lc->sc->rpl;
1406+ const int min_pu_width = fc->ps.pps->min_pu_width;
1407+ const MvField* tab_mvf = fc->tab.mvf;
1408+ const MvField *mvf = &TAB_MVF(x_cand, y_cand);
1409+ const PredFlag maskx = lx + 1;
1410+ const int poc = rpl[lx].list[ref_idx[lx]];
1411+ int available = 0;
1412+
1413+ if ((mvf->pred_flag & maskx) && rpl[lx].list[mvf->ref_idx[lx]] == poc) {
1414+ available = 1;
1415+ *mv = mvf->mv[lx];
1416+ } else {
1417+ const int ly = !lx;
1418+ const PredFlag masky = ly + 1;
1419+ if ((mvf->pred_flag & masky) && rpl[ly].list[mvf->ref_idx[ly]] == poc) {
1420+ available = 1;
1421+ *mv = mvf->mv[ly];
1422+ }
1423+ }
1424+
1425+ return available;
1426+}
1427+
1428+static int affine_mvp_candidate(const VVCLocalContext *lc,
1429+ const int x_cand, const int y_cand, const int lx, const int8_t *ref_idx,
1430+ Mv *cps, const int num_cp)
1431+{
1432+ const VVCFrameContext *fc = lc->fc;
1433+ int x_nb, y_nb, nbw, nbh, motion_model_idc, available = 0;
1434+
1435+ motion_model_idc = affine_neighbour_cb(fc, x_cand, y_cand, &x_nb, &y_nb, &nbw, &nbh);
1436+ if (motion_model_idc) {
1437+ const int min_pu_width = fc->ps.pps->min_pu_width;
1438+ const MvField* tab_mvf = fc->tab.mvf;
1439+ const MvField *mvf = &TAB_MVF(x_nb, y_nb);
1440+ RefPicList* rpl = lc->sc->rpl;
1441+ const PredFlag maskx = lx + 1;
1442+ const int poc = rpl[lx].list[ref_idx[lx]];
1443+
1444+ if ((mvf->pred_flag & maskx) && rpl[lx].list[mvf->ref_idx[lx]] == poc) {
1445+ available = 1;
1446+ affine_cps_from_nb(lc, x_nb, y_nb, nbw, nbh, lx, cps, num_cp);
1447+ } else {
1448+ const int ly = !lx;
1449+ const PredFlag masky = ly + 1;
1450+ if ((mvf->pred_flag & masky) && rpl[ly].list[mvf->ref_idx[ly]] == poc) {
1451+ available = 1;
1452+ affine_cps_from_nb(lc, x_nb, y_nb, nbw, nbh, ly, cps, num_cp);
1453+ }
1454+ }
1455+
1456+ }
1457+ return available;
1458+}
1459+
1460+static int mvp_from_nbs(NeighbourContext *ctx,
1461+ const NeighbourIdx *nbs, const int num_nbs, const int lx, const int8_t *ref_idx, const int amvr_shift,
1462+ Mv *cps, const int num_cps)
1463+{
1464+ const VVCLocalContext *lc = ctx->lc;
1465+ const int is_mvp = 1;
1466+ int available = 0;
1467+
1468+ for (int i = 0; i < num_nbs; i++) {
1469+ Neighbour *n = &ctx->neighbours[nbs[i]];
1470+ if (check_available(n, lc, is_mvp)) {
1471+ if (num_cps > 1)
1472+ available = affine_mvp_candidate(lc, n->x, n->y, lx, ref_idx, cps, num_cps);
1473+ else
1474+ available = mvp_candidate(lc, n->x, n->y, lx, ref_idx, cps);
1475+ if (available) {
1476+ for (int c = 0; c < num_cps; c++)
1477+ ff_vvc_round_mv(cps + c, amvr_shift, amvr_shift);
1478+ return 1;
1479+ }
1480+ }
1481+ }
1482+ return 0;
1483+}
1484+
1485+//get mvp from neighbours
1486+#define AFFINE_MVP_FROM_NBS(nbs) \
1487+ mvp_from_nbs(&nctx, nbs, FF_ARRAY_ELEMS(nbs), lx, ref_idx, amvr_shift, cps, num_cp) \
1488+
1489+#define MVP_FROM_NBS(nbs) \
1490+ mvp_from_nbs(&nctx, nbs, FF_ARRAY_ELEMS(nbs), lx, ref_idx, amvr_shift, mv, 1) \
1491+
1492+static int mvp_spatial_candidates(const VVCLocalContext *lc,
1493+ const int mvp_lx_flag, const int lx, const int8_t* ref_idx, const int amvr_shift,
1494+ Mv* mv, int *nb_merge_cand)
1495+{
1496+ const NeighbourIdx ak[] = { A0, A1 };
1497+ const NeighbourIdx bk[] = { B0, B1, B2 };
1498+ NeighbourContext nctx;
1499+ int available_a, num_cands = 0;
1500+ Mv mv_a;
1501+
1502+ init_neighbour_context(&nctx, lc);
1503+
1504+ available_a = MVP_FROM_NBS(ak);
1505+ if (available_a) {
1506+ if (mvp_lx_flag == num_cands)
1507+ return 1;
1508+ num_cands++;
1509+ mv_a = *mv;
1510+ }
1511+ if (MVP_FROM_NBS(bk)) {
1512+ if (!available_a || !IS_SAME_MV(&mv_a, mv)) {
1513+ if (mvp_lx_flag == num_cands)
1514+ return 1;
1515+ num_cands++;
1516+ }
1517+ }
1518+ *nb_merge_cand = num_cands;
1519+ return 0;
1520+}
1521+
1522+static int mvp_temporal_candidates(const VVCLocalContext* lc,
1523+ const int mvp_lx_flag, const int lx, const int8_t *ref_idx, const int amvr_shift,
1524+ Mv* mv, int *num_cands)
1525+{
1526+ if (temporal_luma_motion_vector(lc, ref_idx[lx], mv, lx, 1, 0)) {
1527+ if (mvp_lx_flag == *num_cands) {
1528+ ff_vvc_round_mv(mv, amvr_shift, amvr_shift);
1529+ return 1;
1530+ }
1531+ (*num_cands)++;
1532+ }
1533+ return 0;
1534+
1535+}
1536+
1537+static int mvp_history_candidates(const VVCLocalContext *lc,
1538+ const int mvp_lx_flag, const int lx, const int8_t ref_idx, const int amvr_shift,
1539+ Mv *mv, int num_cands)
1540+{
1541+ const EntryPoint* ep = lc->ep;
1542+ const RefPicList* rpl = lc->sc->rpl;
1543+ const int poc = rpl[lx].list[ref_idx];
1544+
1545+ if (ep->num_hmvp == 0)
1546+ return 0;
1547+ for (int i = 1; i <= FFMIN(4, ep->num_hmvp); i++) {
1548+ const MvField* h = &ep->hmvp[i - 1];
1549+ for (int j = 0; j < 2; j++) {
1550+ const int ly = (j ? !lx : lx);
1551+ PredFlag mask = PF_L0 + ly;
1552+ if ((h->pred_flag & mask) && poc == rpl[ly].list[h->ref_idx[ly]]) {
1553+ if (mvp_lx_flag == num_cands) {
1554+ *mv = h->mv[ly];
1555+ ff_vvc_round_mv(mv, amvr_shift, amvr_shift);
1556+ return 1;
1557+ }
1558+ num_cands++;
1559+ }
1560+ }
1561+ }
1562+ return 0;
1563+}
1564+
1565+//8.5.2.8 Derivation process for luma motion vector prediction
1566+static void mvp(const VVCLocalContext *lc, const int mvp_lx_flag, const int lx,
1567+ const int8_t *ref_idx, const int amvr_shift, Mv *mv)
1568+{
1569+ int num_cands;
1570+
1571+ if (mvp_spatial_candidates(lc, mvp_lx_flag, lx, ref_idx, amvr_shift, mv, &num_cands))
1572+ return;
1573+
1574+ if (mvp_temporal_candidates(lc, mvp_lx_flag, lx, ref_idx, amvr_shift, mv, &num_cands))
1575+ return;
1576+
1577+ if (mvp_history_candidates(lc, mvp_lx_flag, lx, ref_idx[lx], amvr_shift, mv, num_cands))
1578+ return;
1579+
1580+ memset(mv, 0, sizeof(*mv));
1581+}
1582+
1583+void ff_vvc_mvp(VVCLocalContext *lc, const int *mvp_lx_flag, const int amvr_shift, MotionInfo *mi)
1584+{
1585+ const CodingUnit *cu = lc->cu;
1586+ mi->num_sb_x = 1;
1587+ mi->num_sb_y = 1;
1588+
1589+ ff_vvc_set_neighbour_available(lc, cu->x0, cu->y0, cu->cb_width, cu->cb_height);
1590+ if (mi->pred_flag != PF_L1)
1591+ mvp(lc, mvp_lx_flag[L0], L0, mi->ref_idx, amvr_shift, &mi->mv[L0][0]);
1592+ if (mi->pred_flag != PF_L0)
1593+ mvp(lc, mvp_lx_flag[L1], L1, mi->ref_idx, amvr_shift, &mi->mv[L1][0]);
1594+}
1595+
1596+static int affine_mvp_constructed_cp(NeighbourContext *ctx,
1597+ const NeighbourIdx *neighbour, const int num_neighbour,
1598+ const int lx, const int8_t ref_idx, const int amvr_shift, Mv *cp)
1599+{
1600+ const VVCLocalContext *lc = ctx->lc;
1601+ const VVCFrameContext *fc = lc->fc;
1602+ const MvField *tab_mvf = fc->tab.mvf;
1603+ const int min_pu_width = fc->ps.pps->min_pu_width;
1604+ const RefPicList* rpl = lc->sc->rpl;
1605+ const int is_mvp = 1;
1606+ int available = 0;
1607+
1608+ for (int i = 0; i < num_neighbour; i++) {
1609+ Neighbour *n = &ctx->neighbours[neighbour[i]];
1610+ if (check_available(n, ctx->lc, is_mvp)) {
1611+ const PredFlag maskx = lx + 1;
1612+ const MvField* mvf = &TAB_MVF(n->x, n->y);
1613+ const int poc = rpl[lx].list[ref_idx];
1614+ if ((mvf->pred_flag & maskx) && rpl[lx].list[mvf->ref_idx[lx]] == poc) {
1615+ available = 1;
1616+ *cp = mvf->mv[lx];
1617+ } else {
1618+ const int ly = !lx;
1619+ const PredFlag masky = ly + 1;
1620+ if ((mvf->pred_flag & masky) && rpl[ly].list[mvf->ref_idx[ly]] == poc) {
1621+ available = 1;
1622+ *cp = mvf->mv[ly];
1623+ }
1624+ }
1625+ if (available) {
1626+ ff_vvc_round_mv(cp, amvr_shift, amvr_shift);
1627+ return 1;
1628+ }
1629+ }
1630+ }
1631+ return 0;
1632+}
1633+
1634+#define AFFINE_MVP_CONSTRUCTED_CP(cands, cp) \
1635+ affine_mvp_constructed_cp(nctx, cands, FF_ARRAY_ELEMS(cands), lx, ref_idx, \
1636+ amvr_shift, cp)
1637+
1638+//8.5.5.8 Derivation process for constructed affine control point motion vector prediction candidates
1639+static int affine_mvp_const1(NeighbourContext* nctx,
1640+ const int lx, const int8_t ref_idx, const int amvr_shift,
1641+ Mv *cps, int *available)
1642+{
1643+ const NeighbourIdx tl[] = { B2, B3, A2 };
1644+ const NeighbourIdx tr[] = { B1, B0 };
1645+ const NeighbourIdx bl[] = { A1, A0 };
1646+
1647+ available[0] = AFFINE_MVP_CONSTRUCTED_CP(tl, cps + 0);
1648+ available[1] = AFFINE_MVP_CONSTRUCTED_CP(tr, cps + 1);
1649+ available[2] = AFFINE_MVP_CONSTRUCTED_CP(bl, cps + 2);
1650+ return available[0] && available[1];
1651+}
1652+
1653+//8.5.5.7 item 7
1654+static void affine_mvp_const2(const int idx, Mv *cps, const int num_cp)
1655+{
1656+ const Mv mv = cps[idx];
1657+ for (int j = 0; j < num_cp; j++)
1658+ cps[j] = mv;
1659+}
1660+
1661+//8.5.5.7 Derivation process for luma affine control point motion vector predictors
1662+static void affine_mvp(const VVCLocalContext *lc,
1663+ const int mvp_lx_flag, const int lx, const int8_t *ref_idx, const int amvr_shift,
1664+ MotionModelIdc motion_model_idc, Mv *cps)
1665+{
1666+ const NeighbourIdx ak[] = { A0, A1 };
1667+ const NeighbourIdx bk[] = { B0, B1, B2 };
1668+ const int num_cp = motion_model_idc + 1;
1669+ NeighbourContext nctx;
1670+ int available[MAX_CONTROL_POINTS];
1671+ int num_cands = 0;
1672+
1673+ init_neighbour_context(&nctx, lc);
1674+ //Ak
1675+ if (AFFINE_MVP_FROM_NBS(ak)) {
1676+ if (mvp_lx_flag == num_cands)
1677+ return;
1678+ num_cands++;
1679+ }
1680+ //Bk
1681+ if (AFFINE_MVP_FROM_NBS(bk)) {
1682+ if (mvp_lx_flag == num_cands)
1683+ return;
1684+ num_cands++;
1685+ }
1686+
1687+ //Const1
1688+ if (affine_mvp_const1(&nctx, lx, ref_idx[lx], amvr_shift, cps, available)) {
1689+ if (available[2] || motion_model_idc == MOTION_4_PARAMS_AFFINE) {
1690+ if (mvp_lx_flag == num_cands)
1691+ return;
1692+ num_cands++;
1693+ }
1694+ }
1695+
1696+ //Const2
1697+ for (int i = 2; i >= 0; i--) {
1698+ if (available[i]) {
1699+ if (mvp_lx_flag == num_cands) {
1700+ affine_mvp_const2(i, cps, num_cp);
1701+ return;
1702+ }
1703+ num_cands++;
1704+ }
1705+ }
1706+ if (temporal_luma_motion_vector(lc, ref_idx[lx], cps, lx, 1, 0)) {
1707+ if (mvp_lx_flag == num_cands) {
1708+ ff_vvc_round_mv(cps, amvr_shift, amvr_shift);
1709+ for (int i = 1; i < num_cp; i++)
1710+ cps[i] = cps[0];
1711+ return;
1712+ }
1713+ num_cands++;
1714+ }
1715+
1716+ //Zero Mv
1717+ memset(cps, 0, num_cp * sizeof(Mv));
1718+}
1719+
1720+void ff_vvc_affine_mvp(VVCLocalContext *lc, const int *mvp_lx_flag, const int amvr_shift, MotionInfo *mi)
1721+{
1722+ const CodingUnit *cu = lc->cu;
1723+
1724+ mi->num_sb_x = cu->cb_width >> MIN_PU_LOG2;
1725+ mi->num_sb_y = cu->cb_height >> MIN_PU_LOG2;
1726+
1727+ ff_vvc_set_neighbour_available(lc, cu->x0, cu->y0, cu->cb_width, cu->cb_height);
1728+ if (mi->pred_flag != PF_L1)
1729+ affine_mvp(lc, mvp_lx_flag[L0], L0, mi->ref_idx, amvr_shift, mi->motion_model_idc, &mi->mv[L0][0]);
1730+ if (mi->pred_flag != PF_L0)
1731+ affine_mvp(lc, mvp_lx_flag[L1], L1, mi->ref_idx, amvr_shift, mi->motion_model_idc, &mi->mv[L1][0]);
1732+}
1733+
1734+//8.5.2.14 Rounding process for motion vectors
1735+void ff_vvc_round_mv(Mv *mv, const int lshift, const int rshift)
1736+{
1737+ if (rshift) {
1738+ const int offset = 1 << (rshift - 1);
1739+ mv->x = ((mv->x + offset - (mv->x >= 0)) >> rshift) << lshift;
1740+ mv->y = ((mv->y + offset - (mv->y >= 0)) >> rshift) << lshift;
1741+ } else {
1742+ mv->x = mv->x << lshift;
1743+ mv->y = mv->y << lshift;
1744+ }
1745+}
1746+
1747+void ff_vvc_clip_mv(Mv *mv)
1748+{
1749+ mv->x = av_clip(mv->x, -(1 << 17), (1 << 17) - 1);
1750+ mv->y = av_clip(mv->y, -(1 << 17), (1 << 17) - 1);
1751+}
1752+
1753+//8.5.2.1 Derivation process for motion vector components and reference indices
1754+static av_always_inline int is_greater_mer(const VVCFrameContext *fc, const int x0, const int y0, const int x0_br, const int y0_br)
1755+{
1756+ const uint8_t plevel = fc->ps.sps->log2_parallel_merge_level;
1757+
1758+ return x0_br >> plevel > x0 >> plevel &&
1759+ y0_br >> plevel > y0 >> plevel;
1760+}
1761+
1762+//8.5.2.16 Updating process for the history-based motion vector predictor candidate list
1763+void ff_vvc_update_hmvp(VVCLocalContext *lc, const MotionInfo *mi)
1764+{
1765+ const VVCFrameContext *fc = lc->fc;
1766+ const CodingUnit *cu = lc->cu;
1767+ const int min_pu_width = fc->ps.pps->min_pu_width;
1768+ const MvField* tab_mvf = fc->tab.mvf;
1769+ EntryPoint* ep = lc->ep;
1770+ const MvField *mvf;
1771+ int i;
1772+
1773+ if (!is_greater_mer(fc, cu->x0, cu->y0, cu->x0 + cu->cb_width, cu->y0 + cu->cb_height))
1774+ return;
1775+ mvf = &TAB_MVF(cu->x0, cu->y0);
1776+
1777+ for (i = 0; i < ep->num_hmvp; i++) {
1778+ if (compare_mv_ref_idx(mvf, ep->hmvp + i)) {
1779+ ep->num_hmvp--;
1780+ break;
1781+ }
1782+ }
1783+ if (i == MAX_NUM_HMVP_CANDS) {
1784+ ep->num_hmvp--;
1785+ i = 0;
1786+ }
1787+
1788+ memmove(ep->hmvp + i, ep->hmvp + i + 1, (ep->num_hmvp - i) * sizeof(MvField));
1789+ ep->hmvp[ep->num_hmvp++] = *mvf;
1790+}
1791+
1792+MvField* ff_vvc_get_mvf(const VVCFrameContext *fc, const int x0, const int y0)
1793+{
1794+ const int min_pu_width = fc->ps.pps->min_pu_width;
1795+ MvField* tab_mvf = fc->tab.mvf;
1796+
1797+ return &TAB_MVF(x0, y0);
1798+}
libavcodec/vvc/vvc_mvs.h+46−0
@@ -0,0 +1,46 @@
1+/*
2+ * VVC motion vector decoder
3+ *
4+ * Copyright (C) 2023 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+
23+#ifndef AVCODEC_VVC_VVC_MVS_H
24+#define AVCODEC_VVC_VVC_MVS_H
25+
26+#include "vvcdec.h"
27+
28+void ff_vvc_round_mv(Mv *mv, int lshift, int rshift);
29+void ff_vvc_clip_mv(Mv *mv);
30+void ff_vvc_mv_scale(Mv *dst, const Mv *src, int td, int tb);
31+void ff_vvc_luma_mv_merge_mode(VVCLocalContext *lc, int merge_idx, int ciip_flag, MvField *mv);
32+void ff_vvc_luma_mv_merge_gpm(VVCLocalContext *lc, const int merge_gpm_idx[2], MvField *mv);
33+void ff_vvc_mvp(VVCLocalContext *lc, const int *mvp_lx_flag, const int amvr_shift, MotionInfo *mi);
34+void ff_vvc_sb_mv_merge_mode(VVCLocalContext *lc, int merge_subblock_idx, PredictionUnit *pu);
35+void ff_vvc_affine_mvp(VVCLocalContext *lc, const int *mvp_lx_flag, const int amvr_shift, MotionInfo* mi);
36+void ff_vvc_store_sb_mvs(const VVCLocalContext *lc, PredictionUnit *pu);
37+void ff_vvc_store_mv(const VVCLocalContext *lc, const MotionInfo *mi);
38+void ff_vvc_store_mvf(const VVCLocalContext *lc, const MvField *mvf);
39+void ff_vvc_store_gpm_mvf(const VVCLocalContext *lc, const PredictionUnit* pu);
40+void ff_vvc_update_hmvp(VVCLocalContext *lc, const MotionInfo *mi);
41+int ff_vvc_no_backward_pred_flag(const VVCLocalContext *lc);
42+MvField* ff_vvc_get_mvf(const VVCFrameContext *fc, const int x0, const int y0);
43+void ff_vvc_set_mvf(const VVCLocalContext *lc, const int x0, const int y0, const int w, const int h, const MvField *mvf);
44+void ff_vvc_set_intra_mvf(const VVCLocalContext *lc);
45+
46+#endif //AVCODEC_VVC_VVC_MVS_H

Keyboard shortcuts

General

Show keyboard shortcuts
?
Focus the current search
/

Issues and pull requests

Open filters
f

Issues, pull requests and Inbox

Select next item
j↓
Select previous item
k↑
Open selected item
Enter
Dismiss filters or clear selection
Esc

Repository code

Go to file
t