diff mbox series

[FFmpeg-devel,06/14] vvcdec: add motion vector decoder

Message ID 20230521130319.13813-7-nuomi2021@gmail.com
State Superseded
Headers show
Series add vvc decoder c code | expand

Checks

Context Check Description
andriy/make_x86 fail Make failed

Commit Message

Nuo Mi May 21, 2023, 1:03 p.m. UTC
---
 libavcodec/vvc/Makefile  |    4 +-
 libavcodec/vvc/vvc_ctu.c |   41 +
 libavcodec/vvc/vvc_ctu.h |    4 +
 libavcodec/vvc/vvc_mvs.c | 1807 ++++++++++++++++++++++++++++++++++++++
 libavcodec/vvc/vvc_mvs.h |   46 +
 5 files changed, 1901 insertions(+), 1 deletion(-)
 create mode 100644 libavcodec/vvc/vvc_ctu.c
 create mode 100644 libavcodec/vvc/vvc_mvs.c
 create mode 100644 libavcodec/vvc/vvc_mvs.h
diff mbox series

Patch

diff --git a/libavcodec/vvc/Makefile b/libavcodec/vvc/Makefile
index 14c23a6fdf..7c59b99f15 100644
--- a/libavcodec/vvc/Makefile
+++ b/libavcodec/vvc/Makefile
@@ -6,4 +6,6 @@  OBJS-$(CONFIG_VVC_DECODER)          +=  vvc/vvcdec.o            \
                                         vvc/vvc_ps.o            \
                                         vvc/vvc_data.o          \
                                         vvc/vvc_cabac.o         \
-                                        vvc/vvc_refs.o
+                                        vvc/vvc_refs.o          \
+                                        vvc/vvc_mvs.o           \
+                                        vvc/vvc_ctu.o
diff --git a/libavcodec/vvc/vvc_ctu.c b/libavcodec/vvc/vvc_ctu.c
new file mode 100644
index 0000000000..4e6e582718
--- /dev/null
+++ b/libavcodec/vvc/vvc_ctu.c
@@ -0,0 +1,41 @@ 
+/*
+ * VVC CTU(Coding Tree Unit) parser
+ *
+ * Copyright (C) 2022 Nuo Mi
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "vvc_cabac.h"
+#include "vvc_ctu.h"
+#include "vvc_data.h"
+#include "vvc_mvs.h"
+
+void ff_vvc_set_neighbour_available(VVCLocalContext *lc,
+    const int x0, const int y0, const int w, const int h)
+{
+    const int log2_ctb_size = lc->fc->ps.sps->ctb_log2_size_y;
+    const int x0b = av_mod_uintp2(x0, log2_ctb_size);
+    const int y0b = av_mod_uintp2(y0, log2_ctb_size);
+
+    lc->na.cand_up       = (lc->ctb_up_flag   || y0b);
+    lc->na.cand_left     = (lc->ctb_left_flag || x0b);
+    lc->na.cand_up_left  = (x0b || y0b) ? lc->na.cand_left && lc->na.cand_up : lc->ctb_up_left_flag;
+    lc->na.cand_up_right_sap =
+            (x0b + w == 1 << log2_ctb_size) ? lc->ctb_up_right_flag && !y0b : lc->na.cand_up;
+    lc->na.cand_up_right = lc->na.cand_up_right_sap && (x0 + w) < lc->end_of_tiles_x;
+}
diff --git a/libavcodec/vvc/vvc_ctu.h b/libavcodec/vvc/vvc_ctu.h
index 7da9e15b45..02c757559e 100644
--- a/libavcodec/vvc/vvc_ctu.h
+++ b/libavcodec/vvc/vvc_ctu.h
@@ -23,6 +23,8 @@ 
 #ifndef AVCODEC_VVC_CTU_H
 #define AVCODEC_VVC_CTU_H
 
+#include "libavcodec/cabac.h"
+
 #include "vvcdec.h"
 
 enum SAOType {
@@ -384,4 +386,6 @@  struct ALFParams {
     uint8_t applied[3];
 };
 
+//utils
+void ff_vvc_set_neighbour_available(VVCLocalContext *lc, int x0, int y0, int w, int h);
 #endif // AVCODEC_VVC_CTU_H
diff --git a/libavcodec/vvc/vvc_mvs.c b/libavcodec/vvc/vvc_mvs.c
new file mode 100644
index 0000000000..39261ad26e
--- /dev/null
+++ b/libavcodec/vvc/vvc_mvs.c
@@ -0,0 +1,1807 @@ 
+/*
+ * VVC motion vector decoder
+ *
+ * Copyright (C) 2023 Nuo Mi
+ * Copyright (C) 2022 Xu Mu
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "vvc_ctu.h"
+#include "vvc_data.h"
+#include "vvc_refs.h"
+#include "vvc_mvs.h"
+
+#define IS_SAME_MV(a, b) (AV_RN64A(a) == AV_RN64A(b))
+
+//check if the two luma locations belong to the same motion estimation region
+static av_always_inline int is_same_mer(const VVCFrameContext *fc, const int xN, const int yN, const int xP, const int yP)
+{
+    const uint8_t plevel = fc->ps.sps->log2_parallel_merge_level;
+
+    return xN >> plevel == xP >> plevel &&
+           yN >> plevel == yP >> plevel;
+}
+
+//return true if we have same mvs and ref_idxs
+static av_always_inline int compare_mv_ref_idx(const MvField *n, const MvField *o)
+{
+    if (!o || n->pred_flag != o->pred_flag)
+        return 0;
+    for (int i = 0; i < 2; i++) {
+        PredFlag mask = i + 1;
+        if (n->pred_flag & mask) {
+            const int same_ref_idx = n->ref_idx[i] == o->ref_idx[i];
+            const int same_mv = IS_SAME_MV(n->mv + i, o->mv + i);
+            if (!same_ref_idx || !same_mv)
+                return 0;
+        }
+    }
+    return 1;
+}
+
+// 8.5.2.15 Temporal motion buffer compression process for collocated motion vectors
+static av_always_inline void mv_compression(Mv *motion)
+{
+    int mv[2] = {motion->x, motion->y};
+    for (int i = 0; i < 2; i++) {
+        const int s = mv[i] >> 17;
+        const int f = av_log2((mv[i] ^ s) | 31) - 4;
+        const int mask  = (-1 << f) >> 1;
+        const int round = (1 << f) >> 2;
+        mv[i] = (mv[i] + round) & mask;
+    }
+    motion->x = mv[0];
+    motion->y = mv[1];
+}
+
+void ff_vvc_mv_scale(Mv *dst, const Mv *src, int td, int tb)
+{
+    int tx, scale_factor;
+
+    td = av_clip_int8(td);
+    tb = av_clip_int8(tb);
+    tx = (0x4000 + (abs(td) >> 1)) / td;
+    scale_factor = av_clip_intp2((tb * tx + 32) >> 6, 12);
+    dst->x = av_clip_intp2((scale_factor * src->x + 127 +
+                           (scale_factor * src->x < 0)) >> 8, 17);
+    dst->y = av_clip_intp2((scale_factor * src->y + 127 +
+                           (scale_factor * src->y < 0)) >> 8, 17);
+}
+
+//part of 8.5.2.12 Derivation process for collocated motion vectors
+static int check_mvset(Mv *mvLXCol, Mv *mvCol,
+                       int colPic, int poc,
+                       const RefPicList *refPicList, int X, int refIdxLx,
+                       const RefPicList *refPicList_col, int listCol, int refidxCol)
+{
+    int cur_lt = refPicList[X].isLongTerm[refIdxLx];
+    int col_lt = refPicList_col[listCol].isLongTerm[refidxCol];
+    int col_poc_diff, cur_poc_diff;
+
+    if (cur_lt != col_lt) {
+        mvLXCol->x = 0;
+        mvLXCol->y = 0;
+        return 0;
+    }
+
+    col_poc_diff = colPic - refPicList_col[listCol].list[refidxCol];
+    cur_poc_diff = poc    - refPicList[X].list[refIdxLx];
+
+    mv_compression(mvCol);
+    if (cur_lt || col_poc_diff == cur_poc_diff) {
+        mvLXCol->x = av_clip_intp2(mvCol->x, 17);
+        mvLXCol->y = av_clip_intp2(mvCol->y, 17);
+    } else {
+        ff_vvc_mv_scale(mvLXCol, mvCol, col_poc_diff, cur_poc_diff);
+    }
+    return 1;
+}
+
+#define CHECK_MVSET(l)                                          \
+    check_mvset(mvLXCol, temp_col.mv + l,                       \
+                colPic, fc->ps.ph->poc,                         \
+                refPicList, X, refIdxLx,                        \
+                refPicList_col, L ## l, temp_col.ref_idx[l])
+
+//derive NoBackwardPredFlag
+int ff_vvc_no_backward_pred_flag(const VVCFrameContext *fc)
+{
+    int check_diffpicount = 0;
+    int i, j;
+    const RefPicList *refPicList = fc->ref->refPicList;
+
+    for (j = 0; j < 2; j++) {
+        for (i = 0; i < refPicList[j].nb_refs; i++) {
+            if (refPicList[j].list[i] > fc->ps.ph->poc) {
+                check_diffpicount++;
+                break;
+            }
+        }
+    }
+    return !check_diffpicount;
+}
+
+//8.5.2.12 Derivation process for collocated motion vectors
+static int derive_temporal_colocated_mvs(const VVCLocalContext *lc, MvField temp_col,
+                                         int refIdxLx, Mv *mvLXCol, int X,
+                                         int colPic, const RefPicList *refPicList_col, int sb_flag)
+{
+    const VVCFrameContext *fc   = lc->fc;
+    RefPicList* refPicList      = fc->ref->refPicList;
+
+    if (temp_col.pred_flag == PF_INTRA)
+        return 0;
+
+    if (sb_flag){
+        if (X == 0) {
+            if (temp_col.pred_flag & PF_L0)
+                return CHECK_MVSET(0);
+            else if (ff_vvc_no_backward_pred_flag(fc) && (temp_col.pred_flag & PF_L1))
+                return CHECK_MVSET(1);
+        } else {
+            if (temp_col.pred_flag & PF_L1)
+                return CHECK_MVSET(1);
+            else if (ff_vvc_no_backward_pred_flag(fc) && (temp_col.pred_flag & PF_L0))
+                return CHECK_MVSET(0);
+        }
+    } else {
+        if (!(temp_col.pred_flag & PF_L0))
+            return CHECK_MVSET(1);
+        else if (temp_col.pred_flag == PF_L0)
+            return CHECK_MVSET(0);
+        else if (temp_col.pred_flag == PF_BI) {
+            if (ff_vvc_no_backward_pred_flag(fc)) {
+                if (X == 0)
+                    return CHECK_MVSET(0);
+                else
+                    return CHECK_MVSET(1);
+            } else {
+                if (lc->sc->sh.collocated_list)
+                    return CHECK_MVSET(0);
+                else
+                    return CHECK_MVSET(1);
+            }
+        }
+    }
+
+    return 0;
+}
+
+#define TAB_MVF(x, y)                                                   \
+    tab_mvf[((y) >> MIN_PU_LOG2) * min_pu_width + ((x) >> MIN_PU_LOG2)]
+
+#define TAB_MVF_PU(v)                                                   \
+    TAB_MVF(x ## v, y ## v)
+
+#define TAB_CP_MV(lx, x, y)                                              \
+    fc->tab.cp_mv[lx][((((y) >> min_cb_log2_size) * min_cb_width + ((x) >> min_cb_log2_size)) ) * MAX_CONTROL_POINTS]
+
+
+#define DERIVE_TEMPORAL_COLOCATED_MVS(sb_flag)                          \
+    derive_temporal_colocated_mvs(lc, temp_col,                          \
+                                  refIdxLx, mvLXCol, X, colPic,         \
+                                  ff_vvc_get_ref_list(fc, ref, x, y), sb_flag)
+
+//8.5.2.11 Derivation process for temporal luma motion vector prediction
+static int temporal_luma_motion_vector(const VVCLocalContext *lc,
+    const int refIdxLx, Mv *mvLXCol, const int X, int check_center, int sb_flag)
+{
+    const VVCFrameContext *fc   = lc->fc;
+    const VVCSPS *sps           = fc->ps.sps;
+    const CodingUnit *cu        = lc->cu;
+    int x, y, colPic, availableFlagLXCol = 0;
+    int min_pu_width = fc->ps.pps->min_pu_width;
+    VVCFrame *ref = fc->ref->collocated_ref;
+    MvField *tab_mvf;
+    MvField temp_col;
+
+    if (!ref) {
+        memset(mvLXCol, 0, sizeof(*mvLXCol));
+        return 0;
+    }
+
+    if (!fc->ps.ph->temporal_mvp_enabled_flag || (cu->cb_width * cu->cb_height <= 32))
+        return 0;
+
+    tab_mvf = ref->tab_mvf;
+    colPic  = ref->poc;
+
+    //bottom right collocated motion vector
+    x = cu->x0 + cu->cb_width;
+    y = cu->y0 + cu->cb_height;
+
+    if (tab_mvf &&
+        (cu->y0 >> sps->ctb_log2_size_y) == (y >> sps->ctb_log2_size_y) &&
+        y < fc->ps.sps->height &&
+        x < fc->ps.sps->width) {
+        x                 &= ~7;
+        y                 &= ~7;
+        ff_vvc_await_progress(ref, y);
+        temp_col           = TAB_MVF(x, y);
+        availableFlagLXCol = DERIVE_TEMPORAL_COLOCATED_MVS(sb_flag);
+    }
+    if (check_center) {
+        // derive center collocated motion vector
+        if (tab_mvf && !availableFlagLXCol) {
+            x                  = cu->x0 + (cu->cb_width >> 1);
+            y                  = cu->y0 + (cu->cb_height >> 1);
+            x                 &= ~7;
+            y                 &= ~7;
+            ff_vvc_await_progress(ref, y);
+            temp_col           = TAB_MVF(x, y);
+            availableFlagLXCol = DERIVE_TEMPORAL_COLOCATED_MVS(sb_flag);
+        }
+    }
+    return availableFlagLXCol;
+}
+
+void ff_vvc_set_mvf(const VVCLocalContext *lc, const int x0, const int y0, const int w, const int h, const MvField *mvf)
+{
+    const VVCFrameContext *fc   = lc->fc;
+    MvField *tab_mvf            = fc->ref->tab_mvf;
+    const int min_pu_width      = fc->ps.pps->min_pu_width;
+    const int min_pu_size       = 1 << MIN_PU_LOG2;
+    for (int dy = 0; dy < h; dy += min_pu_size) {
+        for (int dx = 0; dx < w; dx += min_pu_size) {
+            const int x = x0 + dx;
+            const int y = y0 + dy;
+            TAB_MVF(x, y) = *mvf;
+        }
+    }
+}
+
+void ff_vvc_set_intra_mvf(const VVCLocalContext *lc)
+{
+    const VVCFrameContext *fc   = lc->fc;
+    const CodingUnit *cu        = lc->cu;
+    MvField *tab_mvf            = fc->ref->tab_mvf;
+    const int min_pu_width      = fc->ps.pps->min_pu_width;
+    const int min_pu_size       = 1 << MIN_PU_LOG2;
+    for (int dy = 0; dy < cu->cb_height; dy += min_pu_size) {
+        for (int dx = 0; dx < cu->cb_width; dx += min_pu_size) {
+            const int x = cu->x0 + dx;
+            const int y = cu->y0 + dy;
+            TAB_MVF(x, y).pred_flag = PF_INTRA;
+        }
+    }
+}
+
+//cbProfFlagLX from 8.5.5.9 Derivation process for motion vector arrays from affine control point motion vectors
+static int derive_cb_prof_flag_lx(const VVCLocalContext *lc, const PredictionUnit* pu, int lx, int is_fallback)
+{
+    const MotionInfo* mi    = &pu->mi;
+    const Mv* cp_mv         = &mi->mv[lx][0];
+    if (lc->fc->ps.ph->prof_disabled_flag || is_fallback)
+        return 0;
+    if (mi->motion_model_idc == MOTION_4_PARAMS_AFFINE) {
+        if (IS_SAME_MV(cp_mv, cp_mv + 1))
+            return 0;
+    }
+    if (mi->motion_model_idc == MOTION_6_PARAMS_AFFINE) {
+        if (IS_SAME_MV(cp_mv, cp_mv + 1) && IS_SAME_MV(cp_mv, cp_mv + 2))
+            return 0;
+    }
+    //fixme: RprConstraintsActiveFlag
+    return 1;
+}
+
+typedef struct SubblockParams {
+    int d_hor_x;
+    int d_ver_x;
+    int d_hor_y;
+    int d_ver_y;
+    int mv_scale_hor;
+    int mv_scale_ver;
+    int is_fallback;
+
+    int cb_width;
+    int cb_height;
+} SubblockParams;
+
+static int is_fallback_mode(const SubblockParams *sp, const PredFlag pred_flag)
+{
+    const int a = 4 * (2048 + sp->d_hor_x);
+    const int b = 4 * sp->d_hor_y;
+    const int c = 4 * (2048 + sp->d_ver_y);
+    const int d = 4 * sp->d_ver_x;
+    if (pred_flag == PF_BI) {
+        const int max_w4 = FFMAX(0, FFMAX(a, FFMAX(b, a + b)));
+        const int min_w4 = FFMIN(0, FFMIN(a, FFMIN(b, a + b)));
+        const int max_h4 = FFMAX(0, FFMAX(c, FFMAX(d, c + d)));
+        const int min_h4 = FFMIN(0, FFMIN(c, FFMIN(d, c + d)));
+        const int bx_wx4 = ((max_w4 - min_w4) >> 11) + 9;
+        const int bx_hx4 = ((max_h4 - min_h4) >> 11) + 9;
+        return bx_wx4 * bx_hx4 > 225;
+    } else {
+        const int bx_wxh = (FFABS(a) >> 11) + 9;
+        const int bx_hxh = (FFABS(d) >> 11) + 9;
+        const int bx_wxv = (FFABS(b) >> 11) + 9;
+        const int bx_hxv = (FFABS(c) >> 11) + 9;
+        if (bx_wxh * bx_hxh <= 165 && bx_wxv * bx_hxv <= 165)
+            return 0;
+    }
+    return 1;
+
+}
+
+static void init_subblock_params(SubblockParams *sp, const MotionInfo* mi,
+    const int cb_width, const int cb_height, const int lx)
+{
+    const int log2_cbw  = av_log2(cb_width);
+    const int log2_cbh  = av_log2(cb_height);
+    const Mv* cp_mv     = mi->mv[lx];
+    const int num_cp_mv = mi->motion_model_idc + 1;
+    sp->d_hor_x = (cp_mv[1].x - cp_mv[0].x) << (MAX_CU_DEPTH - log2_cbw);
+    sp->d_ver_x = (cp_mv[1].y - cp_mv[0].y) << (MAX_CU_DEPTH - log2_cbw);
+    if (num_cp_mv == 3) {
+        sp->d_hor_y = (cp_mv[2].x - cp_mv[0].x) << (MAX_CU_DEPTH - log2_cbh);
+        sp->d_ver_y = (cp_mv[2].y - cp_mv[0].y) << (MAX_CU_DEPTH - log2_cbh);
+    } else {
+        sp->d_hor_y = -sp->d_ver_x;
+        sp->d_ver_y = sp->d_hor_x;
+    }
+    sp->mv_scale_hor = (cp_mv[0].x) << MAX_CU_DEPTH;
+    sp->mv_scale_ver = (cp_mv[0].y) << MAX_CU_DEPTH;
+    sp->cb_width  = cb_width;
+    sp->cb_height = cb_height;
+    sp->is_fallback = is_fallback_mode(sp, mi->pred_flag);
+}
+
+static void derive_subblock_diff_mvs(const VVCLocalContext *lc, PredictionUnit* pu, const SubblockParams* sp, const int lx)
+{
+    pu->cb_prof_flag[lx] = derive_cb_prof_flag_lx(lc, pu, lx, sp->is_fallback);
+    if (pu->cb_prof_flag[lx]) {
+        const int dmv_limit = 1 << 5;
+        const int pos_offset_x = 6 * (sp->d_hor_x + sp->d_hor_y);
+        const int pos_offset_y = 6 * (sp->d_ver_x + sp->d_ver_y);
+        for (int x = 0; x < AFFINE_MIN_BLOCK_SIZE; x++) {
+            for (int y = 0; y < AFFINE_MIN_BLOCK_SIZE; y++) {
+                Mv diff;
+                diff.x = x * (sp->d_hor_x << 2) + y * (sp->d_hor_y << 2) - pos_offset_x;
+                diff.y = x * (sp->d_ver_x << 2) + y * (sp->d_ver_y << 2) - pos_offset_y;
+                ff_vvc_round_mv(&diff, 0, 8);
+                pu->diff_mv_x[lx][AFFINE_MIN_BLOCK_SIZE * y + x] = av_clip(diff.x, -dmv_limit + 1, dmv_limit - 1);
+                pu->diff_mv_y[lx][AFFINE_MIN_BLOCK_SIZE * y + x] = av_clip(diff.y, -dmv_limit + 1, dmv_limit - 1);
+            }
+        }
+    }
+}
+
+static void store_cp_mv(const VVCLocalContext *lc, const MotionInfo *mi, const int lx)
+{
+    VVCFrameContext *fc = lc->fc;
+    const CodingUnit *cu = lc->cu;
+    const int log2_min_cb_size = fc->ps.sps->min_cb_log2_size_y;
+    const int min_cb_size = fc->ps.sps->min_cb_size_y;
+    const int min_cb_width = fc->ps.pps->min_cb_width;
+    const int num_cp_mv = mi->motion_model_idc + 1;
+
+    for (int dy = 0; dy < cu->cb_height; dy += min_cb_size) {
+        for (int dx = 0; dx < cu->cb_width; dx += min_cb_size) {
+            const int x_cb = (cu->x0 + dx) >> log2_min_cb_size;
+            const int y_cb = (cu->y0 + dy) >> log2_min_cb_size;
+            const int offset = (y_cb * min_cb_width + x_cb) * MAX_CONTROL_POINTS;
+
+            memcpy(&fc->tab.cp_mv[lx][offset], mi->mv[lx], sizeof(Mv) * num_cp_mv);
+            SAMPLE_CTB(fc->tab.mmi, x_cb, y_cb) = mi->motion_model_idc;
+        }
+    }
+}
+
+//8.5.5.9 Derivation process for motion vector arrays from affine control point motion vectors
+void ff_vvc_store_sb_mvs(const VVCLocalContext *lc, PredictionUnit *pu)
+{
+    const CodingUnit *cu = lc->cu;
+    const MotionInfo *mi = &pu->mi;
+    const int sbw = cu->cb_width / mi->num_sb_x;
+    const int sbh = cu->cb_height / mi->num_sb_y;
+    SubblockParams params[2];
+    MvField mvf;
+
+    mvf.pred_flag = mi->pred_flag;
+    mvf.bcw_idx = mi->bcw_idx;
+    mvf.hpel_if_idx = mi->hpel_if_idx;
+    mvf.ciip_flag = 0;
+    for (int i = 0; i < 2; i++) {
+        const PredFlag mask = i + 1;
+        if (mi->pred_flag & mask) {
+            store_cp_mv(lc, mi, i);
+            init_subblock_params(params + i, mi, cu->cb_width, cu->cb_height, i);
+            derive_subblock_diff_mvs(lc, pu, params + i, i);
+            mvf.ref_idx[i] = mi->ref_idx[i];
+        }
+    }
+
+    for (int sby = 0; sby < mi->num_sb_y; sby++) {
+        for (int sbx = 0; sbx < mi->num_sb_x; sbx++) {
+            const int x0 = cu->x0 + sbx * sbw;
+            const int y0 = cu->y0 + sby * sbh;
+            for (int i = 0; i < 2; i++) {
+                const PredFlag mask = i + 1;
+                if (mi->pred_flag & mask) {
+                    const SubblockParams* sp = params + i;
+                    const int x_pos_cb = sp->is_fallback ? (cu->cb_width >> 1) : (2 + (sbx << MIN_CU_LOG2));
+                    const int y_pos_cb = sp->is_fallback ? (cu->cb_height >> 1) : (2 + (sby << MIN_CU_LOG2));
+                    Mv *mv = mvf.mv + i;
+
+                    mv->x = sp->mv_scale_hor + sp->d_hor_x * x_pos_cb + sp->d_hor_y * y_pos_cb;
+                    mv->y = sp->mv_scale_ver + sp->d_ver_x * x_pos_cb + sp->d_ver_y * y_pos_cb;
+                    ff_vvc_round_mv(mv, 0, MAX_CU_DEPTH);
+                    ff_vvc_clip_mv(mv);
+                }
+            }
+            ff_vvc_set_mvf(lc, x0, y0, sbw, sbh, &mvf);
+        }
+    }
+}
+
+void ff_vvc_store_gpm_mvf(const VVCLocalContext *lc, const PredictionUnit *pu)
+{
+    const CodingUnit *cu     = lc->cu;
+    const int angle_idx      = ff_vvc_gpm_angle_idx[pu->gpm_partition_idx];
+    const int distance_idx   = ff_vvc_gpm_distance_idx[pu->gpm_partition_idx];
+    const int displacement_x = ff_vvc_gpm_distance_lut[angle_idx];
+    const int displacement_y = ff_vvc_gpm_distance_lut[(angle_idx + 8) % 32];
+    const int is_flip        = angle_idx >= 13 &&angle_idx <= 27;
+    const int shift_hor      = (angle_idx % 16 == 8 || (angle_idx % 16 && cu->cb_height >= cu->cb_width)) ? 0 : 1;
+    const int sign           = angle_idx < 16 ? 1 : -1;
+    const int block_size     = 4;
+    int offset_x = (-cu->cb_width) >> 1;
+    int offset_y = (-cu->cb_height) >> 1;
+
+    if (!shift_hor)
+        offset_y += sign * ((distance_idx * cu->cb_height) >> 3);
+    else
+        offset_x += sign * ((distance_idx * cu->cb_width) >> 3);
+
+    for (int y = 0; y < cu->cb_height; y += block_size) {
+        for (int x = 0; x < cu->cb_width; x += block_size) {
+            const int motion_idx = (((x + offset_x) << 1) + 5) * displacement_x +
+                (((y + offset_y) << 1) + 5) * displacement_y;
+            const int s_type = FFABS(motion_idx) < 32 ? 2 : (motion_idx <= 0 ? (1 - is_flip) : is_flip);
+            const int pred_flag = pu->gpm_mv[0].pred_flag | pu->gpm_mv[1].pred_flag;
+            const int x0 = cu->x0 + x;
+            const int y0 = cu->y0 + y;
+
+            if (!s_type)
+                ff_vvc_set_mvf(lc, x0, y0, block_size, block_size, pu->gpm_mv + 0);
+            else if (s_type == 1 || (s_type == 2 && pred_flag != PF_BI))
+                ff_vvc_set_mvf(lc, x0, y0, block_size, block_size, pu->gpm_mv + 1);
+            else {
+                MvField mvf  = pu->gpm_mv[0];
+                const MvField *mv1 = &pu->gpm_mv[1];
+                const int lx =  mv1->pred_flag - PF_L0;
+                mvf.pred_flag = PF_BI;
+                mvf.ref_idx[lx] = mv1->ref_idx[lx];
+                mvf.mv[lx] = mv1->mv[lx];
+                ff_vvc_set_mvf(lc, x0, y0, block_size, block_size, &mvf);
+            }
+        }
+    }
+
+}
+
+void ff_vvc_store_mvf(const VVCLocalContext *lc, const MvField *mvf)
+{
+    const CodingUnit *cu = lc->cu;
+    ff_vvc_set_mvf(lc, cu->x0, cu->y0, cu->cb_width, cu->cb_height, mvf);
+}
+
+void ff_vvc_store_mv(const VVCLocalContext *lc, const MotionInfo *mi)
+{
+    const CodingUnit *cu = lc->cu;
+    MvField mvf;
+
+    mvf.hpel_if_idx = mi->hpel_if_idx;
+    mvf.bcw_idx = mi->bcw_idx;
+    mvf.pred_flag = mi->pred_flag;
+    mvf.ciip_flag = 0;
+
+    for (int i = 0; i < 2; i++) {
+        const PredFlag mask = i + 1;
+        if (mvf.pred_flag & mask) {
+            mvf.mv[i] = mi->mv[i][0];
+            mvf.ref_idx[i] = mi->ref_idx[i];
+        }
+    }
+    ff_vvc_set_mvf(lc, cu->x0, cu->y0, cu->cb_width, cu->cb_height, &mvf);
+}
+
+typedef enum NeighbourIdx {
+    A0,
+    A1,
+    A2,
+    B0,
+    B1,
+    B2,
+    B3,
+    NUM_NBS,
+    NB_IDX_NONE = NUM_NBS,
+} NeighbourIdx;
+
+typedef struct Neighbour {
+    int x;
+    int y;
+
+    int checked;
+    int available;
+} Neighbour;
+
+typedef struct NeighbourContext {
+    Neighbour neighbours[NUM_NBS];
+    const VVCLocalContext *lc;
+} NeighbourContext;
+
+static int is_a0_available(const VVCLocalContext *lc, const CodingUnit *cu)
+{
+    const VVCFrameContext *fc   = lc->fc;
+    const VVCSPS *sps           = fc->ps.sps;
+    const int x0b               = av_mod_uintp2(cu->x0, sps->ctb_log2_size_y);
+    int cand_bottom_left;
+
+    if (!x0b && !lc->ctb_left_flag) {
+        cand_bottom_left = 0;
+    } else {
+        const int log2_min_cb_size  = sps->min_cb_log2_size_y;
+        const int min_cb_width      = fc->ps.pps->min_cb_width;
+        const int x                 = (cu->x0 - 1) >> log2_min_cb_size;
+        const int y                 = (cu->y0 + cu->cb_height) >> log2_min_cb_size;
+        const int max_y             = FFMIN(fc->ps.pps->height, ((cu->y0 >> sps->ctb_log2_size_y) + 1) << sps->ctb_log2_size_y);
+        if (cu->y0 + cu->cb_height >= max_y)
+            cand_bottom_left = 0;
+        else
+            cand_bottom_left = SAMPLE_CTB(fc->tab.cb_width[0], x, y) != 0;
+    }
+    return cand_bottom_left;
+}
+
+static void init_neighbour_context(NeighbourContext *ctx, const VVCLocalContext *lc)
+{
+    const CodingUnit *cu            = lc->cu;
+    const NeighbourAvailable *na    = &lc->na;
+    const int x0                    = cu->x0;
+    const int y0                    = cu->y0;
+    const int cb_width              = cu->cb_width;
+    const int cb_height             = cu->cb_height;
+    const int a0_available          = is_a0_available(lc, cu);
+
+    Neighbour neighbours[NUM_NBS] = {
+        { x0 - 1,               y0 + cb_height,         !a0_available           }, //A0
+        { x0 - 1,               y0 + cb_height - 1,     !na->cand_left          }, //A1
+        { x0 - 1,               y0,                     !na->cand_left          }, //A2
+        { x0 + cb_width,        y0 - 1,                 !na->cand_up_right      }, //B0
+        { x0 + cb_width - 1,    y0 - 1,                 !na->cand_up            }, //B1
+        { x0 - 1,               y0 - 1,                 !na->cand_up_left       }, //B2
+        { x0,                   y0 - 1,                 !na->cand_up            }, //B3
+    };
+
+    memcpy(ctx->neighbours, neighbours, sizeof(neighbours));
+    ctx->lc = lc;
+}
+
+static int check_available(Neighbour *n, const VVCLocalContext *lc, const int is_mvp)
+{
+    const VVCFrameContext *fc   = lc->fc;
+    const VVCSPS *sps           = fc->ps.sps;
+    const CodingUnit *cu        = lc->cu;
+    const MvField *tab_mvf      = fc->ref->tab_mvf;
+    const int min_pu_width      = fc->ps.pps->min_pu_width;
+
+    if (!n->checked) {
+        n->checked = 1;
+        n->available = !sps->entropy_coding_sync_enabled_flag || ((n->x >> sps->ctb_log2_size_y) <= (cu->x0 >> sps->ctb_log2_size_y));
+        n->available &= TAB_MVF(n->x, n->y).pred_flag != PF_INTRA;
+        if (!is_mvp)
+            n->available &= !is_same_mer(fc, n->x, n->y, cu->x0, cu->y0);
+    }
+    return n->available;
+}
+
+static const MvField *mv_merge_candidate(const VVCLocalContext *lc, const int x_cand, const int y_cand)
+{
+    const VVCFrameContext *fc   = lc->fc;
+    const int min_pu_width      = fc->ps.pps->min_pu_width;
+    const MvField* tab_mvf      = fc->ref->tab_mvf;
+    const MvField *mvf          = &TAB_MVF(x_cand, y_cand);
+
+    return mvf;
+}
+
+static const MvField* mv_merge_from_nb(NeighbourContext *ctx, const NeighbourIdx nb)
+{
+    const VVCLocalContext *lc   = ctx->lc;
+    const int is_mvp            = 0;
+    Neighbour *n                = &ctx->neighbours[nb];
+
+    if (check_available(n, lc, is_mvp))
+        return mv_merge_candidate(lc, n->x, n->y);
+    return 0;
+}
+#define MV_MERGE_FROM_NB(nb) mv_merge_from_nb(&nctx, nb)
+
+//8.5.2.3 Derivation process for spatial merging candidates
+static int mv_merge_spatial_candidates(const VVCLocalContext *lc, const int merge_idx,
+    const MvField **nb_list, MvField *cand_list, int *nb_merge_cand)
+{
+    const MvField *cand;
+    int num_cands = 0;
+    NeighbourContext nctx;
+
+    static NeighbourIdx nbs[][2] = {
+        {B1, NB_IDX_NONE },
+        {A1, B1 },
+        {B0, B1 },
+        {A0, A1 },
+    };
+
+    init_neighbour_context(&nctx, lc);
+    for (int i = 0; i < FF_ARRAY_ELEMS(nbs); i++) {
+        NeighbourIdx nb    = nbs[i][0];
+        NeighbourIdx old   = nbs[i][1];
+        cand = nb_list[nb] = MV_MERGE_FROM_NB(nb);
+        if (cand && !compare_mv_ref_idx(cand, nb_list[old])) {
+            cand_list[num_cands] = *cand;
+            if (merge_idx == num_cands)
+                return 1;
+            num_cands++;
+        }
+    }
+    if (num_cands != 4) {
+        cand = MV_MERGE_FROM_NB(B2);
+        if (cand && !compare_mv_ref_idx(cand, nb_list[A1])
+            && !compare_mv_ref_idx(cand, nb_list[B1])) {
+            cand_list[num_cands] = *cand;
+            if (merge_idx == num_cands)
+                return 1;
+            num_cands++;
+        }
+    }
+    *nb_merge_cand = num_cands;
+    return 0;
+}
+
+static int mv_merge_temporal_candidate(const VVCLocalContext *lc, MvField *cand)
+{
+    const VVCFrameContext *fc   = lc->fc;
+    const CodingUnit *cu        = lc->cu;
+    memset(cand, 0, sizeof(*cand));
+    if (fc->ps.ph->temporal_mvp_enabled_flag && (cu->cb_width * cu->cb_height > 32)) {
+        int available_l0 = temporal_luma_motion_vector(lc, 0, cand->mv + 0, 0, 1, 0);
+        int available_l1 = (lc->sc->sh.slice_type == VVC_SLICE_TYPE_B) ?
+            temporal_luma_motion_vector(lc, 0, cand->mv + 1, 1, 1, 0) : 0;
+        cand->pred_flag = available_l0 + (available_l1 << 1);
+    }
+    return cand->pred_flag;
+}
+
+//8.5.2.6 Derivation process for history-based merging candidates
+static int mv_merge_history_candidates(const VVCLocalContext *lc, const int merge_idx,
+    const MvField **nb_list, MvField *cand_list, int *num_cands)
+{
+    const VVCSPS *sps       = lc->fc->ps.sps;
+    const EntryPoint* ep    = lc->ep;
+    for (int i = 1; i <= ep->num_hmvp && (*num_cands < sps->max_num_merge_cand - 1); i++) {
+        const MvField *h = &ep->hmvp[ep->num_hmvp - i];
+        const int same_motion = i <= 2 && (compare_mv_ref_idx(h, nb_list[A1]) || compare_mv_ref_idx(h, nb_list[B1]));
+        if (!same_motion) {
+            cand_list[*num_cands] = *h;
+            if (merge_idx == *num_cands)
+                return 1;
+            (*num_cands)++;
+        }
+    }
+    return 0;
+}
+
+//8.5.2.4 Derivation process for pairwise average merging candidate
+static int mv_merge_pairwise_candidate(MvField *cand_list, const int num_cands, const int is_b)
+{
+    if (num_cands > 1) {
+        const int num_ref_rists = is_b ? 2 : 1;
+        const MvField* p0       = cand_list + 0;
+        const MvField* p1       = cand_list + 1;
+        MvField* cand           = cand_list + num_cands;
+
+        cand->pred_flag = 0;
+        for (int i = 0; i < num_ref_rists; i++) {
+            PredFlag mask = i + 1;
+            if (p0->pred_flag & mask) {
+                cand->pred_flag |= mask;
+                cand->ref_idx[i] = p0->ref_idx[i];
+                if (p1->pred_flag & mask) {
+                    Mv *mv = cand->mv + i;
+                    mv->x = p0->mv[i].x + p1->mv[i].x;
+                    mv->y = p0->mv[i].y + p1->mv[i].y;
+                    ff_vvc_round_mv(mv, 0, 1);
+                } else {
+                    cand->mv[i] = p0->mv[i];
+                }
+            } else if (p1->pred_flag & mask) {
+                cand->pred_flag |= mask;
+                cand->mv[i] = p1->mv[i];
+                cand->ref_idx[i] = p1->ref_idx[i];
+            }
+        }
+        if (cand->pred_flag) {
+            cand->hpel_if_idx = p0->hpel_if_idx == p1->hpel_if_idx ? p0->hpel_if_idx : 0;
+            cand->bcw_idx = 0;
+            cand->ciip_flag = 0;
+            return 1;
+        }
+    }
+    return 0;
+}
+
+//8.5.2.5 Derivation process for zero motion vector merging candidates
+static void mv_merge_zero_motion_candidate(const VVCLocalContext *lc, const int merge_idx,
+    MvField *cand_list, int num_cands)
+{
+    const VVCSPS *sps   = lc->fc->ps.sps;
+    const VVCSH *sh     = &lc->sc->sh;
+    const int nb_refs   = (sh->slice_type == VVC_SLICE_TYPE_P) ?
+        sh->rpls[0].num_ref_entries : FFMIN(sh->rpls[0].num_ref_entries, sh->rpls[1].num_ref_entries);
+    int zero_idx        = 0;
+
+    while (num_cands < sps->max_num_merge_cand) {
+        MvField *cand = cand_list + num_cands;
+
+        cand->pred_flag    = PF_L0 + ((sh->slice_type == VVC_SLICE_TYPE_B) << 1);
+        AV_ZERO64(cand->mv + 0);
+        AV_ZERO64(cand->mv + 1);
+        cand->ref_idx[0]   = zero_idx < nb_refs ? zero_idx : 0;
+        cand->ref_idx[1]   = zero_idx < nb_refs ? zero_idx : 0;
+        cand->bcw_idx      = 0;
+        cand->hpel_if_idx  = 0;
+        if (merge_idx == num_cands)
+            return;
+        num_cands++;
+        zero_idx++;
+    }
+}
+
+static void mv_merge_mode(const VVCLocalContext *lc,  const int merge_idx, MvField *cand_list)
+{
+    int num_cands    = 0;
+    const MvField *nb_list[NUM_NBS + 1] = { NULL };
+
+    if (mv_merge_spatial_candidates(lc, merge_idx, nb_list, cand_list, &num_cands))
+        return;
+
+    if (mv_merge_temporal_candidate(lc, &cand_list[num_cands])) {
+        if (merge_idx == num_cands)
+            return;
+        num_cands++;
+    }
+
+    if (mv_merge_history_candidates(lc, merge_idx, nb_list, cand_list, &num_cands))
+        return;
+
+    if (mv_merge_pairwise_candidate(cand_list, num_cands, lc->sc->sh.slice_type == VVC_SLICE_TYPE_B)) {
+        if (merge_idx == num_cands)
+            return;
+        num_cands++;
+    }
+
+    mv_merge_zero_motion_candidate(lc, merge_idx, cand_list, num_cands);
+}
+
+//8.5.2.2 Derivation process for luma motion vectors for merge mode
+void ff_vvc_luma_mv_merge_mode(VVCLocalContext *lc, const int merge_idx, const int ciip_flag, MvField *mv)
+{
+    const CodingUnit *cu = lc->cu;
+    MvField cand_list[MRG_MAX_NUM_CANDS];
+
+    ff_vvc_set_neighbour_available(lc, cu->x0, cu->y0, cu->cb_width, cu->cb_height);
+    mv_merge_mode(lc, merge_idx, cand_list);
+    *mv = cand_list[merge_idx];
+    //ciip flag in not inhritable
+    mv->ciip_flag = ciip_flag;
+}
+
+//8.5.4.2 Derivation process for luma motion vectors for geometric partitioning merge mode
+void ff_vvc_luma_mv_merge_gpm(VVCLocalContext *lc, const int merge_gpm_idx[2], MvField *mv)
+{
+    const CodingUnit *cu = lc->cu;
+    MvField cand_list[MRG_MAX_NUM_CANDS];
+
+    const int idx[] = { merge_gpm_idx[0], merge_gpm_idx[1] + (merge_gpm_idx[1] >= merge_gpm_idx[0]) };
+
+    ff_vvc_set_neighbour_available(lc, cu->x0, cu->y0, cu->cb_width, cu->cb_height);
+    mv_merge_mode(lc, FFMAX(idx[0], idx[1]), cand_list);
+    memset(mv, 0, 2 * sizeof(*mv));
+    for (int i = 0; i < 2; i++) {
+        int lx   = idx[i] & 1;
+        int mask = lx + PF_L0;
+        MvField *cand = cand_list + idx[i];
+        if (!(cand->pred_flag & mask)) {
+            lx   = !lx;
+            mask = lx + PF_L0;
+        }
+        mv[i].pred_flag   = mask;
+        mv[i].ref_idx[lx] = cand->ref_idx[lx];
+        mv[i].mv[lx]      = cand->mv[lx];
+    }
+
+}
+
+//8.5.5.5 Derivation process for luma affine control point motion vectors from a neighbouring block
+static void affine_cps_from_nb(const VVCLocalContext *lc,
+    const int x_nb, int y_nb, const int nbw, const int nbh, const int lx,
+    Mv *cps, int num_cps)
+{
+    const VVCFrameContext *fc   = lc->fc;
+    const CodingUnit *cu        = lc->cu;
+    const int x0                = cu->x0;
+    const int y0                = cu->y0;
+    const int cb_width          = cu->cb_width;
+    const int cb_height         = cu->cb_height;
+    const MvField* tab_mvf      = fc->ref->tab_mvf;
+    const int min_cb_log2_size  = fc->ps.sps->min_cb_log2_size_y;
+    const int min_cb_width      = fc->ps.pps->min_cb_width;
+
+    const int log2_nbw          = ff_log2(nbw);
+    const int log2_nbh          = ff_log2(nbh);
+    const int is_ctb_boundary   = !((y_nb + nbh) % fc->ps.sps->ctb_size_y) && (y_nb + nbh == y0);
+    const Mv *l, *r;
+    int mv_scale_hor, mv_scale_ver, d_hor_x, d_ver_x, d_hor_y, d_ver_y, motion_model_idc_nb;
+    if (is_ctb_boundary) {
+        const int min_pu_width = fc->ps.pps->min_pu_width;
+        l = &TAB_MVF(x_nb, y_nb + nbh - 1).mv[lx];
+        r = &TAB_MVF(x_nb + nbw - 1, y_nb + nbh - 1).mv[lx];
+    } else {
+        const int x = x_nb >> min_cb_log2_size;
+        const int y = y_nb >> min_cb_log2_size;
+        motion_model_idc_nb  = SAMPLE_CTB(fc->tab.mmi, x, y);
+
+        l = &TAB_CP_MV(lx, x_nb, y_nb);
+        r = &TAB_CP_MV(lx, x_nb + nbw - 1, y_nb) + 1;
+    }
+    mv_scale_hor = l->x << 7;
+    mv_scale_ver = l->y << 7;
+    d_hor_x = (r->x - l->x) << (7 - log2_nbw);
+    d_ver_x = (r->y - l->y) << (7 - log2_nbw);
+    if (!is_ctb_boundary && motion_model_idc_nb == MOTION_6_PARAMS_AFFINE) {
+        const Mv* lb = &TAB_CP_MV(lx, x_nb, y_nb + nbh - 1) + 2;
+        d_hor_y = (lb->x - l->x) << (7 - log2_nbh);
+        d_ver_y = (lb->y - l->y) << (7 - log2_nbh);
+    } else {
+        d_hor_y = -d_ver_x;
+        d_ver_y = d_hor_x;
+    }
+
+    if (is_ctb_boundary) {
+        y_nb = y0;
+    }
+    cps[0].x = mv_scale_hor + d_hor_x * (x0 - x_nb)  + d_hor_y * (y0 - y_nb);
+    cps[0].y = mv_scale_ver + d_ver_x * (x0 - x_nb)  + d_ver_y * (y0 - y_nb);
+    cps[1].x = mv_scale_hor + d_hor_x * (x0 + cb_width - x_nb)  + d_hor_y * (y0 - y_nb);
+    cps[1].y = mv_scale_ver + d_ver_x * (x0 + cb_width - x_nb)  + d_ver_y * (y0 - y_nb);
+    if (num_cps == 3) {
+        cps[2].x = mv_scale_hor + d_hor_x * (x0 - x_nb)  + d_hor_y * (y0 + cb_height - y_nb);
+        cps[2].y = mv_scale_ver + d_ver_x * (x0 - x_nb)  + d_ver_y * (y0 + cb_height - y_nb);
+    }
+    for (int i = 0; i < num_cps; i++) {
+        ff_vvc_round_mv(cps + i, 0, 7);
+        ff_vvc_clip_mv(cps + i);
+    }
+}
+
+//derive affine neighbour's postion, width and height,
+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)
+{
+    const int log2_min_cb_size  = fc->ps.sps->min_cb_log2_size_y;
+    const int min_cb_width      = fc->ps.pps->min_cb_width;
+    const int x                 = x_nb >> log2_min_cb_size;
+    const int y                 = y_nb >> log2_min_cb_size;
+    const int motion_model_idc  = SAMPLE_CTB(fc->tab.mmi, x, y);
+    if (motion_model_idc) {
+        *x_cb = SAMPLE_CTB(fc->tab.cb_pos_x[0],  x, y);
+        *y_cb = SAMPLE_CTB(fc->tab.cb_pos_y[0],  x, y);
+        *cbw  = SAMPLE_CTB(fc->tab.cb_width[0],  x, y);
+        *cbh  = SAMPLE_CTB(fc->tab.cb_height[0], x, y);
+    }
+    return motion_model_idc;
+}
+
+//part of 8.5.5.2 Derivation process for motion vectors and reference indices in subblock merge mode
+static int affine_merge_candidate(const VVCLocalContext *lc, const int x_cand, const int y_cand, MotionInfo* mi)
+{
+    const VVCFrameContext *fc = lc->fc;
+    int x, y, w, h, motion_model_idc;
+
+    motion_model_idc = affine_neighbour_cb(fc, x_cand, y_cand, &x, &y, &w, &h);
+    if (motion_model_idc) {
+        const int min_pu_width = fc->ps.pps->min_pu_width;
+        const MvField* tab_mvf = fc->ref->tab_mvf;
+        const MvField *mvf = &TAB_MVF(x, y);
+
+        mi->bcw_idx   = mvf->bcw_idx;
+        mi->pred_flag = mvf->pred_flag;
+        for (int i = 0; i < 2; i++) {
+            PredFlag mask = i + 1;
+            if (mi->pred_flag & mask) {
+                affine_cps_from_nb(lc, x, y, w, h, i, &mi->mv[i][0], motion_model_idc + 1);
+            }
+            mi->ref_idx[i] = mvf->ref_idx[i];
+        }
+        mi->motion_model_idc = motion_model_idc;
+    }
+    return motion_model_idc;
+}
+
+static int affine_merge_from_nbs(NeighbourContext *ctx, const NeighbourIdx *nbs, const int num_nbs, MotionInfo* cand)
+{
+    const VVCLocalContext *lc = ctx->lc;
+    const int is_mvp          = 0;
+    for (int i = 0; i < num_nbs; i++) {
+        Neighbour *n = &ctx->neighbours[nbs[i]];
+        if (check_available(n, lc, is_mvp) && affine_merge_candidate(lc, n->x, n->y, cand))
+            return 1;
+    }
+    return 0;
+}
+#define AFFINE_MERGE_FROM_NBS(nbs) affine_merge_from_nbs(&nctx, nbs, FF_ARRAY_ELEMS(nbs), mi)
+
+
+static const MvField* derive_corner_mvf(NeighbourContext *ctx, const NeighbourIdx *neighbour, const int num_neighbour)
+{
+    const VVCFrameContext *fc   = ctx->lc->fc;
+    const MvField *tab_mvf      = fc->ref->tab_mvf;
+    const int min_pu_width      = fc->ps.pps->min_pu_width;
+    for (int i = 0; i < num_neighbour; i++) {
+        Neighbour *n = &ctx->neighbours[neighbour[i]];
+        if (check_available(n, ctx->lc, 0)) {
+            return &TAB_MVF(n->x, n->y);
+        }
+    }
+    return NULL;
+}
+
+#define DERIVE_CORNER_MV(nbs) derive_corner_mvf(nctx, nbs, FF_ARRAY_ELEMS(nbs))
+
+// check if the mv's and refidx are the same between A and B
+static av_always_inline int compare_pf_ref_idx(const MvField *A, const struct MvField *B, const struct MvField *C, const int lx)
+{
+
+    const PredFlag mask = (lx + 1) & A->pred_flag;
+    if (!(B->pred_flag & mask))
+        return 0;
+    if (A->ref_idx[lx] != B->ref_idx[lx])
+        return 0;
+    if (C) {
+        if (!(C->pred_flag & mask))
+            return 0;
+        if (A->ref_idx[lx] != C->ref_idx[lx])
+            return 0;
+    }
+    return 1;
+}
+
+static av_always_inline void sb_clip_location(const VVCFrameContext *fc,
+    const int x_ctb, const int y_ctb, const Mv* temp_mv, int *x, int *y)
+{
+    const VVCPPS *pps       = fc->ps.pps;
+    const int ctb_log2_size = fc->ps.sps->ctb_log2_size_y;
+    *y = av_clip(*y + temp_mv->y, y_ctb, FFMIN(pps->height - 1, y_ctb + (1 << ctb_log2_size) - 1)) & ~7;
+    *x = av_clip(*x + temp_mv->x, x_ctb, FFMIN(pps->width - 1,  x_ctb + (1 << ctb_log2_size) + 3)) & ~7;
+}
+
+static void sb_temproal_luma_motion(const VVCLocalContext *lc,
+    const int x_ctb, const int y_ctb, const Mv *temp_mv,
+    int x, int y, uint8_t *pred_flag, Mv *mv)
+{
+    MvField temp_col;
+    Mv* mvLXCol;
+    const int refIdxLx          = 0;
+    const VVCFrameContext *fc   = lc->fc;
+    const VVCSH *sh             = &lc->sc->sh;
+    const int min_pu_width      = fc->ps.pps->min_pu_width;
+    VVCFrame *ref               = fc->ref->collocated_ref;
+    MvField *tab_mvf            = ref->tab_mvf;
+    int colPic                  = ref->poc;
+    int X                       = 0;
+
+    sb_clip_location(fc, x_ctb, y_ctb, temp_mv, &x, &y);
+
+    ff_vvc_await_progress(ref, y);
+
+    temp_col    = TAB_MVF(x, y);
+    mvLXCol     = mv + 0;
+    *pred_flag = DERIVE_TEMPORAL_COLOCATED_MVS(1);
+    if (sh->slice_type == VVC_SLICE_TYPE_B) {
+        X = 1;
+        mvLXCol = mv + 1;
+        *pred_flag |= (DERIVE_TEMPORAL_COLOCATED_MVS(1)) << 1;
+    }
+}
+
+//8.5.5.4 Derivation process for subblock-based temporal merging base motion data
+static int sb_temporal_luma_motion_data(const VVCLocalContext *lc, const MvField *a1,
+    const int x_ctb, const int y_ctb, MvField *ctr_mvf, Mv *temp_mv)
+{
+    const VVCFrameContext *fc   = lc->fc;
+    const CodingUnit *cu        = lc->cu;
+    const int x                 = cu->x0  + cu->cb_width / 2;
+    const int y                 = cu->y0  + cu->cb_height / 2;
+    const VVCFrame *ref         = fc->ref->collocated_ref;
+
+    int colPic;
+
+    memset(temp_mv, 0, sizeof(*temp_mv));
+
+    if (!ref) {
+        memset(ctr_mvf, 0, sizeof(*ctr_mvf));
+        return 0;
+    }
+
+    colPic  = ref->poc;
+
+    AV_ZERO64(temp_mv);
+    if (a1) {
+        if ((a1->pred_flag & PF_L0) && colPic == fc->ref->refPicList[0].list[a1->ref_idx[0]])
+            *temp_mv = a1->mv[0];
+        else if ((a1->pred_flag & PF_L1) && colPic == fc->ref->refPicList[1].list[a1->ref_idx[1]])
+            *temp_mv = a1->mv[1];
+        ff_vvc_round_mv(temp_mv, 0, 4);
+    }
+    sb_temproal_luma_motion(lc, x_ctb, y_ctb, temp_mv, x, y, &ctr_mvf->pred_flag , ctr_mvf->mv);
+
+    return ctr_mvf->pred_flag;
+}
+
+
+//8.5.5.3 Derivation process for subblock-based temporal merging candidates
+static int sb_temporal_merge_candidate(const VVCLocalContext* lc, NeighbourContext *nctx, PredictionUnit *pu)
+{
+    const VVCFrameContext *fc   = lc->fc;
+    const CodingUnit *cu        = lc->cu;
+    const VVCSPS *sps           = fc->ps.sps;
+    const VVCPH *ph             = fc->ps.ph;
+    MotionInfo *mi              = &pu->mi;
+    const int ctb_log2_size     = sps->ctb_log2_size_y;
+    const int x0                = cu->x0;
+    const int y0                = cu->y0;
+    const NeighbourIdx n        = A1;
+    const MvField *a1;
+    MvField ctr_mvf;
+    Mv temp_mv;
+    const int x_ctb = (x0 >> ctb_log2_size) << ctb_log2_size;
+    const int y_ctb = (y0 >> ctb_log2_size) << ctb_log2_size;
+
+
+    if (!ph->temporal_mvp_enabled_flag ||
+        !sps->sbtmvp_enabled_flag ||
+        (cu->cb_width < 8 && cu->cb_height < 8))
+        return 0;
+
+    mi->num_sb_x = cu->cb_width >> 3;
+    mi->num_sb_y = cu->cb_height >> 3;
+
+    a1 = derive_corner_mvf(nctx, &n, 1);
+    if (sb_temporal_luma_motion_data(lc, a1, x_ctb, y_ctb, &ctr_mvf, &temp_mv)) {
+        const int sbw = cu->cb_width / mi->num_sb_x;
+        const int sbh = cu->cb_height / mi->num_sb_y;
+        MvField mvf = {0};
+        for (int sby = 0; sby < mi->num_sb_y; sby++) {
+            for (int sbx = 0; sbx < mi->num_sb_x; sbx++) {
+                int x = x0 + sbx * sbw;
+                int y = y0 + sby * sbh;
+                sb_temproal_luma_motion(lc, x_ctb, y_ctb, &temp_mv, x + sbw / 2, y +  sbh / 2, &mvf.pred_flag, mvf.mv);
+                if (!mvf.pred_flag) {
+                    mvf.pred_flag = ctr_mvf.pred_flag;
+                    memcpy(mvf.mv, ctr_mvf.mv, sizeof(mvf.mv));
+                }
+                ff_vvc_set_mvf(lc, x, y, sbw, sbh, &mvf);
+            }
+        }
+        return 1;
+    }
+
+    return 0;
+}
+
+static int affine_merge_const1(const MvField *c0, const MvField *c1, const MvField *c2, MotionInfo *mi)
+{
+    if (c0 && c1 && c2) {
+        mi->pred_flag = 0;
+        for (int i = 0; i < 2; i++) {
+            PredFlag mask = i + 1;
+            if (compare_pf_ref_idx(c0, c1, c2, i)) {
+                mi->pred_flag |= mask;
+                mi->ref_idx[i] = c0->ref_idx[i];
+                mi->mv[i][0] = c0->mv[i];
+                mi->mv[i][1] = c1->mv[i];
+                mi->mv[i][2] = c2->mv[i];
+            }
+        }
+        if (mi->pred_flag) {
+            if (mi->pred_flag == PF_BI)
+                mi->bcw_idx = c0->bcw_idx;
+            mi->motion_model_idc = MOTION_6_PARAMS_AFFINE;
+            return 1;
+        }
+    }
+    return 0;
+}
+
+static int affine_merge_const2(const MvField *c0, const MvField *c1, const MvField *c3, MotionInfo *mi)
+{
+    if (c0 && c1 && c3) {
+        mi->pred_flag = 0;
+        for (int i = 0; i < 2; i++) {
+            PredFlag mask = i + 1;
+            if (compare_pf_ref_idx(c0, c1, c3, i)) {
+                mi->pred_flag |= mask;
+                mi->ref_idx[i] = c0->ref_idx[i];
+                mi->mv[i][0] = c0->mv[i];
+                mi->mv[i][1] = c1->mv[i];
+                mi->mv[i][2].x = c3->mv[i].x + c0->mv[i].x - c1->mv[i].x;
+                mi->mv[i][2].y = c3->mv[i].y + c0->mv[i].y - c1->mv[i].y;
+                ff_vvc_clip_mv(&mi->mv[i][2]);
+            }
+        }
+        if (mi->pred_flag) {
+            mi->bcw_idx = mi->pred_flag == PF_BI ? c0->bcw_idx : 0;
+            mi->motion_model_idc = MOTION_6_PARAMS_AFFINE;
+            return 1;
+        }
+    }
+    return 0;
+}
+
+static int affine_merge_const3(const MvField *c0, const MvField *c2, const MvField *c3, MotionInfo *mi)
+{
+    if (c0 && c2 && c3) {
+        mi->pred_flag = 0;
+        for (int i = 0; i < 2; i++) {
+            PredFlag mask = i + 1;
+            if (compare_pf_ref_idx(c0, c2, c3, i)) {
+                mi->pred_flag |= mask;
+                mi->ref_idx[i] = c0->ref_idx[i];
+                mi->mv[i][0] = c0->mv[i];
+                mi->mv[i][1].x = c3->mv[i].x + c0->mv[i].x - c2->mv[i].x;
+                mi->mv[i][1].y = c3->mv[i].y + c0->mv[i].y - c2->mv[i].y;
+                ff_vvc_clip_mv(&mi->mv[i][1]);
+                mi->mv[i][2] = c2->mv[i];
+            }
+        }
+        if (mi->pred_flag) {
+            mi->bcw_idx = mi->pred_flag == PF_BI ? c0->bcw_idx : 0;
+            mi->motion_model_idc = MOTION_6_PARAMS_AFFINE;
+            return 1;
+        }
+    }
+    return 0;
+}
+
+static int affine_merge_const4(const MvField *c1, const MvField *c2, const MvField *c3, MotionInfo *mi)
+{
+    if (c1 && c2 && c3) {
+        mi->pred_flag = 0;
+        for (int i = 0; i < 2; i++) {
+            PredFlag mask = i + 1;
+            if (compare_pf_ref_idx(c1, c2, c3, i)) {
+                mi->pred_flag |= mask;
+                mi->ref_idx[i] = c1->ref_idx[i];
+                mi->mv[i][0].x = c1->mv[i].x + c2->mv[i].x - c3->mv[i].x;
+                mi->mv[i][0].y = c1->mv[i].y + c2->mv[i].y - c3->mv[i].y;
+                ff_vvc_clip_mv(&mi->mv[i][0]);
+                mi->mv[i][1] = c1->mv[i];
+                mi->mv[i][2] = c2->mv[i];
+            }
+        }
+        if (mi->pred_flag) {
+            mi->bcw_idx = mi->pred_flag == PF_BI ? c1->bcw_idx : 0;
+            mi->motion_model_idc = MOTION_6_PARAMS_AFFINE;
+            return 1;
+        }
+    }
+    return 0;
+
+}
+
+static int affine_merge_const5(const MvField *c0, const MvField *c1, MotionInfo *mi)
+{
+    if (c0 && c1) {
+        mi->pred_flag = 0;
+        for (int i = 0; i < 2; i++) {
+            PredFlag mask = i + 1;
+            if (compare_pf_ref_idx(c0, c1, NULL, i)) {
+                mi->pred_flag |= mask;
+                mi->ref_idx[i] = c0->ref_idx[i];
+                mi->mv[i][0] = c0->mv[i];
+                mi->mv[i][1] = c1->mv[i];
+            }
+        }
+        if (mi->pred_flag) {
+            if (mi->pred_flag == PF_BI)
+                mi->bcw_idx = c0->bcw_idx;
+            mi->motion_model_idc = MOTION_4_PARAMS_AFFINE;
+            return 1;
+        }
+    }
+    return 0;
+}
+
+static int affine_merge_const6(const MvField* c0, const MvField* c2, const int cb_width, const int cb_height, MotionInfo *mi)
+{
+    if (c0 && c2) {
+        const int shift = 7 + av_log2(cb_width) - av_log2(cb_height);
+        mi->pred_flag = 0;
+        for (int i = 0; i < 2; i++) {
+            PredFlag mask = i + 1;
+            if (compare_pf_ref_idx(c0, c2, NULL, i)) {
+                mi->pred_flag |= mask;
+                mi->ref_idx[i] = c0->ref_idx[i];
+                mi->mv[i][0] = c0->mv[i];
+                mi->mv[i][1].x = (c0->mv[i].x << 7) + ((c2->mv[i].y - c0->mv[i].y) << shift);
+                mi->mv[i][1].y = (c0->mv[i].y << 7) - ((c2->mv[i].x - c0->mv[i].x) << shift);
+                ff_vvc_round_mv(&mi->mv[i][1], 0, 7);
+                ff_vvc_clip_mv(&mi->mv[i][1]);
+            }
+        }
+        if (mi->pred_flag) {
+            if (mi->pred_flag == PF_BI)
+                mi->bcw_idx = c0->bcw_idx;
+            mi->motion_model_idc = MOTION_4_PARAMS_AFFINE;
+            return 1;
+        }
+    }
+    return 0;
+}
+
+static void affine_merge_zero_motion(const VVCLocalContext *lc, MotionInfo *mi)
+{
+    const CodingUnit *cu = lc->cu;
+    memset(mi, 0, sizeof(*mi));
+    mi->pred_flag    = PF_L0 + ((lc->sc->sh.slice_type == VVC_SLICE_TYPE_B) << 1);
+    mi->motion_model_idc = MOTION_4_PARAMS_AFFINE;
+    mi->num_sb_x = cu->cb_width >> MIN_PU_LOG2;
+    mi->num_sb_y = cu->cb_height >> MIN_PU_LOG2;
+}
+
+//8.5.5.6 Derivation process for constructed affine control point motion vector merging candidates
+static int affine_merge_const_candidates(const VVCLocalContext *lc, MotionInfo *mi,
+    NeighbourContext *nctx, const int merge_subblock_idx, int num_cands)
+{
+    const VVCFrameContext *fc   = lc->fc;
+    const CodingUnit *cu        = lc->cu;
+    const NeighbourIdx tl[]     = { B2, B3, A2 };
+    const NeighbourIdx tr[]     = { B1, B0};
+    const NeighbourIdx bl[]     = { A1, A0};
+    const MvField *c0, *c1, *c2;
+
+    c0 = DERIVE_CORNER_MV(tl);
+    c1 = DERIVE_CORNER_MV(tr);
+    c2 = DERIVE_CORNER_MV(bl);
+
+    if (fc->ps.sps->six_param_affine_enabled_flag) {
+        MvField corner3, *c3 = NULL;
+        //Const1
+        if (affine_merge_const1(c0, c1, c2, mi)) {
+            if (merge_subblock_idx == num_cands)
+                return 1;
+            num_cands++;
+        }
+
+        memset(&corner3, 0, sizeof(corner3));
+        if (fc->ps.ph->temporal_mvp_enabled_flag){
+            const int available_l0 = temporal_luma_motion_vector(lc, 0, corner3.mv + 0, 0, 0, 0);
+            const int available_l1 = (lc->sc->sh.slice_type == VVC_SLICE_TYPE_B) ?
+                temporal_luma_motion_vector(lc, 0, corner3.mv + 1, 1, 0, 0) : 0;
+
+            corner3.pred_flag = available_l0 + (available_l1 << 1);
+            if (corner3.pred_flag)
+                c3 = &corner3;
+        }
+
+        //Const2
+        if (affine_merge_const2(c0, c1, c3, mi)) {
+            if (merge_subblock_idx == num_cands)
+                return 1;
+            num_cands++;
+        }
+
+        //Const3
+        if (affine_merge_const3(c0, c2, c3, mi)) {
+           if (merge_subblock_idx == num_cands)
+               return 1;
+           num_cands++;
+        }
+
+        //Const4
+        if (affine_merge_const4(c1, c2, c3, mi)) {
+           if (merge_subblock_idx == num_cands)
+               return 1;
+           num_cands++;
+        }
+    }
+
+    //Const5
+    if (affine_merge_const5(c0, c1, mi)) {
+        if (merge_subblock_idx == num_cands)
+            return 1;
+        num_cands++;
+    }
+
+    if (affine_merge_const6(c0, c2, cu->cb_width, cu->cb_height, mi)) {
+        if (merge_subblock_idx == num_cands)
+            return 1;
+    }
+    return 0;
+
+}
+
+
+//8.5.5.2 Derivation process for motion vectors and reference indices in subblock merge mode
+//return 1 if candidate is SbCol
+static int sb_mv_merge_mode(const VVCLocalContext *lc, const int merge_subblock_idx, PredictionUnit *pu)
+{
+    const VVCSPS *sps       = lc->fc->ps.sps;
+    const CodingUnit *cu    = lc->cu;
+    MotionInfo *mi          = &pu->mi;
+    int num_cands           = 0;
+    NeighbourContext nctx;
+
+    init_neighbour_context(&nctx, lc);
+
+    //SbCol
+    if (sb_temporal_merge_candidate(lc, &nctx, pu)) {
+        if (merge_subblock_idx == num_cands)
+            return 1;
+        num_cands++;
+    }
+
+    pu->inter_affine_flag = 1;
+    mi->num_sb_x  = cu->cb_width >> MIN_PU_LOG2;
+    mi->num_sb_y  = cu->cb_height >> MIN_PU_LOG2;
+
+    if (sps->affine_enabled_flag) {
+        const NeighbourIdx ak[] = { A0, A1 };
+        const NeighbourIdx bk[] = { B0, B1, B2 };
+        //A
+        if (AFFINE_MERGE_FROM_NBS(ak)) {
+            if (merge_subblock_idx == num_cands)
+                return 0;
+            num_cands++;
+        }
+
+        //B
+        if (AFFINE_MERGE_FROM_NBS(bk)) {
+            if (merge_subblock_idx == num_cands)
+                return 0;
+            num_cands++;
+        }
+
+        //Const1 to Const6
+        if (affine_merge_const_candidates(lc, mi, &nctx, merge_subblock_idx, num_cands))
+            return 0;
+    }
+    //Zero
+    affine_merge_zero_motion(lc, mi);
+    return 0;
+}
+
+void ff_vvc_sb_mv_merge_mode(VVCLocalContext *lc, const int merge_subblock_idx, PredictionUnit *pu)
+{
+    const CodingUnit *cu = lc->cu;
+    ff_vvc_set_neighbour_available(lc, cu->x0, cu->y0, cu->cb_width, cu->cb_height);
+    if (!sb_mv_merge_mode(lc, merge_subblock_idx, pu)) {
+        ff_vvc_store_sb_mvs(lc, pu);
+    }
+}
+
+static int mvp_candidate(const VVCFrameContext *fc,
+    const int x_cand, const int y_cand, const int lx, const int8_t *ref_idx,
+    Mv *mv)
+{
+
+    const int min_pu_width          = fc->ps.pps->min_pu_width;
+    const MvField* tab_mvf          = fc->ref->tab_mvf;
+    const MvField *mvf              = &TAB_MVF(x_cand, y_cand);
+    const RefPicList* refPicList    = fc->ref->refPicList;
+    const PredFlag maskx = lx + 1;
+    const int poc = refPicList[lx].list[ref_idx[lx]];
+    int available = 0;
+
+    if ((mvf->pred_flag & maskx) &&  refPicList[lx].list[mvf->ref_idx[lx]] == poc) {
+        available = 1;
+        *mv = mvf->mv[lx];
+    } else {
+        const int ly = !lx;
+        const PredFlag masky = ly + 1;
+        if ((mvf->pred_flag & masky) &&  refPicList[ly].list[mvf->ref_idx[ly]] == poc) {
+            available = 1;
+            *mv = mvf->mv[ly];
+        }
+    }
+
+    return available;
+}
+
+static int affine_mvp_candidate(const VVCLocalContext *lc,
+    const int x_cand, const int y_cand, const int lx, const int8_t *ref_idx,
+    Mv *cps, const int num_cp)
+{
+    const VVCFrameContext *fc = lc->fc;
+    int x_nb, y_nb, nbw, nbh, motion_model_idc, available = 0;
+
+    motion_model_idc = affine_neighbour_cb(fc, x_cand, y_cand, &x_nb, &y_nb, &nbw, &nbh);
+    if (motion_model_idc) {
+        const int min_pu_width = fc->ps.pps->min_pu_width;
+        const MvField* tab_mvf = fc->ref->tab_mvf;
+        const MvField *mvf = &TAB_MVF(x_nb, y_nb);
+        RefPicList* refPicList = fc->ref->refPicList;
+        const PredFlag maskx = lx + 1;
+        const int poc = refPicList[lx].list[ref_idx[lx]];
+
+        if ((mvf->pred_flag & maskx) &&  refPicList[lx].list[mvf->ref_idx[lx]] == poc) {
+            available = 1;
+            affine_cps_from_nb(lc, x_nb, y_nb, nbw, nbh, lx, cps, num_cp);
+        } else {
+            const int ly = !lx;
+            const PredFlag masky = ly + 1;
+            if ((mvf->pred_flag & masky) &&  refPicList[ly].list[mvf->ref_idx[ly]] == poc) {
+                available = 1;
+                affine_cps_from_nb(lc, x_nb, y_nb, nbw, nbh, ly, cps, num_cp);
+            }
+        }
+
+    }
+    return available;
+}
+
+static int mvp_from_nbs(NeighbourContext *ctx,
+    const NeighbourIdx *nbs, const int num_nbs, const int lx, const int8_t *ref_idx, const int amvr_shift,
+    Mv *cps, const int num_cps)
+{
+    const VVCLocalContext *lc   = ctx->lc;
+    const VVCFrameContext *fc   = lc->fc;
+    const int is_mvp            = 1;
+    int available               = 0;
+
+    for (int i = 0; i < num_nbs; i++) {
+        Neighbour *n = &ctx->neighbours[nbs[i]];
+        if (check_available(n, lc, is_mvp)) {
+            if (num_cps > 1)
+                available = affine_mvp_candidate(lc, n->x, n->y, lx, ref_idx, cps, num_cps);
+            else
+                available = mvp_candidate(fc, n->x, n->y, lx, ref_idx, cps);
+            if (available) {
+                for (int c = 0; c < num_cps; c++)
+                    ff_vvc_round_mv(cps + c, amvr_shift, amvr_shift);
+                return 1;
+            }
+        }
+    }
+    return 0;
+}
+
+//get mvp from neighbours
+#define AFFINE_MVP_FROM_NBS(nbs)                                                         \
+    mvp_from_nbs(&nctx, nbs, FF_ARRAY_ELEMS(nbs), lx, ref_idx, amvr_shift, cps, num_cp)  \
+
+#define MVP_FROM_NBS(nbs)                                                                \
+    mvp_from_nbs(&nctx, nbs, FF_ARRAY_ELEMS(nbs), lx, ref_idx, amvr_shift, mv, 1)        \
+
+static int mvp_spatial_candidates(const VVCLocalContext *lc,
+    const int mvp_lx_flag, const int lx, const int8_t* ref_idx, const int amvr_shift,
+    Mv* mv, int *nb_merge_cand)
+{
+    const NeighbourIdx ak[] = { A0, A1 };
+    const NeighbourIdx bk[] = { B0, B1, B2 };
+    NeighbourContext nctx;
+    int available_a, num_cands = 0;
+    Mv  mv_a;
+
+    init_neighbour_context(&nctx, lc);
+
+    available_a = MVP_FROM_NBS(ak);
+    if (available_a) {
+        if (mvp_lx_flag == num_cands)
+            return 1;
+        num_cands++;
+        mv_a = *mv;
+    }
+    if (MVP_FROM_NBS(bk)) {
+        if (!available_a || !IS_SAME_MV(&mv_a, mv)) {
+            if (mvp_lx_flag == num_cands)
+                return 1;
+            num_cands++;
+        }
+    }
+    *nb_merge_cand = num_cands;
+    return 0;
+}
+
+static int mvp_temporal_candidates(const VVCLocalContext* lc,
+    const int mvp_lx_flag, const int lx, const int8_t *ref_idx, const int amvr_shift,
+    Mv* mv, int *num_cands)
+{
+    if (temporal_luma_motion_vector(lc, ref_idx[lx], mv, lx, 1, 0)) {
+        if (mvp_lx_flag == *num_cands) {
+            ff_vvc_round_mv(mv, amvr_shift, amvr_shift);
+            return 1;
+        }
+        (*num_cands)++;
+    }
+    return 0;
+
+}
+
+static int mvp_history_candidates(const VVCLocalContext *lc,
+    const int mvp_lx_flag, const int lx, const int8_t ref_idx, const int amvr_shift,
+    Mv *mv, int num_cands)
+{
+    const VVCFrameContext *fc       = lc->fc;
+    const EntryPoint* ep            = lc->ep;
+    const RefPicList* refPicList    = fc->ref->refPicList;
+    const int poc                   = refPicList[lx].list[ref_idx];
+
+    if (ep->num_hmvp == 0)
+        return 0;
+    for (int i = 1; i <= FFMIN(4, ep->num_hmvp); i++) {
+        const MvField* h = &ep->hmvp[i - 1];
+        for (int j = 0; j < 2; j++) {
+            const int ly = (j ? !lx : lx);
+            PredFlag mask = PF_L0 + ly;
+            if ((h->pred_flag & mask) && poc == refPicList[ly].list[h->ref_idx[ly]]) {
+                if (mvp_lx_flag == num_cands) {
+                    *mv = h->mv[ly];
+                    ff_vvc_round_mv(mv, amvr_shift, amvr_shift);
+                    return 1;
+                }
+                num_cands++;
+            }
+        }
+    }
+    return 0;
+}
+
+//8.5.2.8 Derivation process for luma motion vector prediction
+static void mvp(const VVCLocalContext *lc, const int mvp_lx_flag, const int lx,
+    const int8_t *ref_idx, const int amvr_shift, Mv *mv)
+{
+    int num_cands;
+
+    if (mvp_spatial_candidates(lc, mvp_lx_flag, lx, ref_idx, amvr_shift, mv, &num_cands))
+        return;
+
+    if (mvp_temporal_candidates(lc, mvp_lx_flag, lx, ref_idx, amvr_shift, mv, &num_cands))
+        return;
+
+    if (mvp_history_candidates(lc, mvp_lx_flag, lx, ref_idx[lx], amvr_shift, mv, num_cands))
+        return;
+
+    memset(mv, 0, sizeof(*mv));
+}
+
+void ff_vvc_mvp(VVCLocalContext *lc, const int *mvp_lx_flag, const int amvr_shift,  MotionInfo *mi)
+{
+    const CodingUnit *cu    = lc->cu;
+    mi->num_sb_x            = 1;
+    mi->num_sb_y            = 1;
+
+    ff_vvc_set_neighbour_available(lc, cu->x0, cu->y0, cu->cb_width, cu->cb_height);
+    if (mi->pred_flag != PF_L1)
+        mvp(lc, mvp_lx_flag[L0], L0, mi->ref_idx, amvr_shift, &mi->mv[L0][0]);
+    if (mi->pred_flag != PF_L0)
+        mvp(lc, mvp_lx_flag[L1], L1, mi->ref_idx, amvr_shift, &mi->mv[L1][0]);
+}
+
+static int affine_mvp_constructed_cp(NeighbourContext *ctx,
+    const NeighbourIdx *neighbour, const int num_neighbour,
+    const int lx, const int8_t ref_idx, const int amvr_shift, Mv *cp)
+{
+    const VVCFrameContext *fc       = ctx->lc->fc;
+    const MvField *tab_mvf          = fc->ref->tab_mvf;
+    const int min_pu_width          = fc->ps.pps->min_pu_width;
+    const RefPicList* refPicList    = fc->ref->refPicList;
+    const int is_mvp                = 1;
+    int available                   = 0;
+
+    for (int i = 0; i < num_neighbour; i++) {
+        Neighbour *n = &ctx->neighbours[neighbour[i]];
+        if (check_available(n, ctx->lc, is_mvp)) {
+            const PredFlag maskx = lx + 1;
+            const MvField* mvf = &TAB_MVF(n->x, n->y);
+            const int poc = refPicList[lx].list[ref_idx];
+            if ((mvf->pred_flag & maskx) &&  refPicList[lx].list[mvf->ref_idx[lx]] == poc) {
+                available = 1;
+                *cp = mvf->mv[lx];
+            } else {
+                const int ly = !lx;
+                const PredFlag masky = ly + 1;
+                if ((mvf->pred_flag & masky) &&  refPicList[ly].list[mvf->ref_idx[ly]] == poc) {
+                    available = 1;
+                    *cp = mvf->mv[ly];
+                }
+            }
+            if (available) {
+                ff_vvc_round_mv(cp, amvr_shift, amvr_shift);
+                return 1;
+            }
+        }
+    }
+    return 0;
+}
+
+#define AFFINE_MVP_CONSTRUCTED_CP(cands, cp)                                    \
+    affine_mvp_constructed_cp(nctx, cands, FF_ARRAY_ELEMS(cands), lx, ref_idx,  \
+        amvr_shift, cp)
+
+//8.5.5.8 Derivation process for constructed affine control point motion vector prediction candidates
+static int affine_mvp_const1(NeighbourContext* nctx,
+    const int lx, const int8_t ref_idx, const int amvr_shift,
+    Mv *cps, int *available)
+{
+    const NeighbourIdx tl[] = { B2, B3, A2 };
+    const NeighbourIdx tr[] = { B1, B0 };
+    const NeighbourIdx bl[] = { A1, A0 };
+
+    available[0] = AFFINE_MVP_CONSTRUCTED_CP(tl, cps + 0);
+    available[1] = AFFINE_MVP_CONSTRUCTED_CP(tr, cps + 1);
+    available[2] = AFFINE_MVP_CONSTRUCTED_CP(bl, cps + 2);
+    return available[0] && available[1];
+}
+
+//8.5.5.7 item 7
+static void affine_mvp_const2(const int idx, Mv *cps, const int num_cp)
+{
+    const Mv mv = cps[idx];
+    for (int j = 0; j < num_cp; j++)
+        cps[j] = mv;
+}
+
+//8.5.5.7 Derivation process for luma affine control point motion vector predictors
+static void affine_mvp(const VVCLocalContext *lc,
+    const int mvp_lx_flag, const int lx, const int8_t *ref_idx, const int amvr_shift,
+    MotionModelIdc motion_model_idc, Mv *cps)
+{
+    const NeighbourIdx ak[] = { A0, A1 };
+    const NeighbourIdx bk[] = { B0, B1, B2 };
+    const int num_cp = motion_model_idc + 1;
+    NeighbourContext nctx;
+    int available[MAX_CONTROL_POINTS];
+    int num_cands    = 0;
+
+    init_neighbour_context(&nctx, lc);
+    //Ak
+    if (AFFINE_MVP_FROM_NBS(ak)) {
+        if (mvp_lx_flag == num_cands)
+            return;
+        num_cands++;
+    }
+    //Bk
+    if (AFFINE_MVP_FROM_NBS(bk)) {
+        if (mvp_lx_flag == num_cands)
+            return;
+        num_cands++;
+    }
+
+    //Const1
+    if (affine_mvp_const1(&nctx, lx, ref_idx[lx], amvr_shift, cps, available)) {
+        if (available[2] || motion_model_idc == MOTION_4_PARAMS_AFFINE) {
+            if (mvp_lx_flag == num_cands)
+                return;
+            num_cands++;
+        }
+    }
+
+    //Const2
+    for (int i = 2; i >= 0; i--) {
+        if (available[i]) {
+            if (mvp_lx_flag == num_cands) {
+                affine_mvp_const2(i, cps, num_cp);
+                return;
+            }
+            num_cands++;
+        }
+    }
+    if (temporal_luma_motion_vector(lc, ref_idx[lx], cps, lx, 1, 0)) {
+        if (mvp_lx_flag == num_cands) {
+            ff_vvc_round_mv(cps, amvr_shift, amvr_shift);
+            for (int i = 1; i < num_cp; i++)
+                cps[i] = cps[0];
+            return;
+        }
+        num_cands++;
+    }
+
+    //Zero Mv
+    memset(cps, 0, num_cp * sizeof(Mv));
+}
+
+void ff_vvc_affine_mvp(VVCLocalContext *lc, const int *mvp_lx_flag, const int amvr_shift,  MotionInfo *mi)
+{
+    const CodingUnit *cu = lc->cu;
+
+    mi->num_sb_x = cu->cb_width >> MIN_PU_LOG2;
+    mi->num_sb_y = cu->cb_height >> MIN_PU_LOG2;
+
+    ff_vvc_set_neighbour_available(lc, cu->x0, cu->y0, cu->cb_width, cu->cb_height);
+    if (mi->pred_flag != PF_L1)
+        affine_mvp(lc, mvp_lx_flag[L0], L0, mi->ref_idx, amvr_shift, mi->motion_model_idc, &mi->mv[L0][0]);
+    if (mi->pred_flag != PF_L0)
+        affine_mvp(lc, mvp_lx_flag[L1], L1, mi->ref_idx, amvr_shift, mi->motion_model_idc, &mi->mv[L1][0]);
+}
+
+//8.5.2.14 Rounding process for motion vectors
+void ff_vvc_round_mv(Mv *mv, const int lshift, const int rshift)
+{
+    if (rshift) {
+        const int offset = 1 << (rshift - 1);
+        mv->x = ((mv->x + offset - (mv->x >= 0)) >> rshift) << lshift;
+        mv->y = ((mv->y + offset - (mv->y >= 0)) >> rshift) << lshift;
+    } else {
+        mv->x = mv->x << lshift;
+        mv->y = mv->y << lshift;
+    }
+}
+
+void ff_vvc_clip_mv(Mv *mv)
+{
+    mv->x = av_clip(mv->x, -(1 << 17), (1 << 17) - 1);
+    mv->y = av_clip(mv->y, -(1 << 17), (1 << 17) - 1);
+}
+
+//8.5.2.1 Derivation process for motion vector components and reference indices
+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)
+{
+    const uint8_t plevel = fc->ps.sps->log2_parallel_merge_level;
+
+    return x0_br >> plevel > x0 >> plevel &&
+           y0_br >> plevel > y0 >> plevel;
+}
+
+//8.5.2.16 Updating process for the history-based motion vector predictor candidate list
+void ff_vvc_update_hmvp(VVCLocalContext *lc, const MotionInfo *mi)
+{
+    const VVCFrameContext *fc   = lc->fc;
+    const CodingUnit *cu        = lc->cu;
+    const int min_pu_width      = fc->ps.pps->min_pu_width;
+    const MvField* tab_mvf      = fc->ref->tab_mvf;
+    EntryPoint* ep              = lc->ep;
+    const MvField *mvf;
+    int i;
+
+    if (!is_greater_mer(fc, cu->x0, cu->y0, cu->x0 + cu->cb_width, cu->y0 + cu->cb_height))
+        return;
+    mvf = &TAB_MVF(cu->x0, cu->y0);
+
+    for (i = 0; i < ep->num_hmvp; i++) {
+        if (compare_mv_ref_idx(mvf, ep->hmvp + i)) {
+            ep->num_hmvp--;
+            break;
+        }
+    }
+    if (i == MAX_NUM_HMVP_CANDS) {
+        ep->num_hmvp--;
+        i = 0;
+    }
+
+    memmove(ep->hmvp + i, ep->hmvp + i + 1, (ep->num_hmvp - i) * sizeof(MvField));
+    ep->hmvp[ep->num_hmvp++] = *mvf;
+}
+
+MvField* ff_vvc_get_mvf(const VVCFrameContext *fc, const int x0, const int y0)
+{
+    const int min_pu_width  = fc->ps.pps->min_pu_width;
+    MvField* tab_mvf        = fc->ref->tab_mvf;
+
+    return &TAB_MVF(x0, y0);
+}
diff --git a/libavcodec/vvc/vvc_mvs.h b/libavcodec/vvc/vvc_mvs.h
new file mode 100644
index 0000000000..4742a18c75
--- /dev/null
+++ b/libavcodec/vvc/vvc_mvs.h
@@ -0,0 +1,46 @@ 
+/*
+ * VVC motion vector decoder
+ *
+ * Copyright (C) 2023 Nuo Mi
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#ifndef AVCODEC_VVC_MVS_H
+#define AVCODEC_VVC_MVS_H
+
+#include "vvcdec.h"
+
+void ff_vvc_round_mv(Mv *mv, int lshift, int rshift);
+void ff_vvc_clip_mv(Mv *mv);
+void ff_vvc_mv_scale(Mv *dst, const Mv *src, int td, int tb);
+void ff_vvc_luma_mv_merge_mode(VVCLocalContext *lc, int merge_idx, int ciip_flag, MvField *mv);
+void ff_vvc_luma_mv_merge_gpm(VVCLocalContext *lc, const int merge_gpm_idx[2], MvField *mv);
+void ff_vvc_mvp(VVCLocalContext *lc, const int *mvp_lx_flag, const int amvr_shift,  MotionInfo *mi);
+void ff_vvc_sb_mv_merge_mode(VVCLocalContext *lc, int merge_subblock_idx, PredictionUnit *pu);
+void ff_vvc_affine_mvp(VVCLocalContext *lc, const int *mvp_lx_flag, const int amvr_shift, MotionInfo* mi);
+void ff_vvc_store_sb_mvs(const VVCLocalContext *lc, PredictionUnit *pu);
+void ff_vvc_store_mv(const VVCLocalContext *lc, const MotionInfo *mi);
+void ff_vvc_store_mvf(const VVCLocalContext *lc, const MvField *mvf);
+void ff_vvc_store_gpm_mvf(const VVCLocalContext *lc, const PredictionUnit* pu);
+void ff_vvc_update_hmvp(VVCLocalContext *lc, const MotionInfo *mi);
+int ff_vvc_no_backward_pred_flag(const VVCFrameContext *fc);
+MvField* ff_vvc_get_mvf(const VVCFrameContext *fc, const int x0, const int y0);
+void ff_vvc_set_mvf(const VVCLocalContext *lc, const int x0, const int y0, const int w, const int h, const MvField *mvf);
+void ff_vvc_set_intra_mvf(const VVCLocalContext *lc);
+
+#endif //AVCODEC_VVC_MVS_H