ref: a46bb83e0c0d9592a20e000e843f36c117c96f13
parent: 64c4cedd3a93549b2a9c490734a76c9a28542934
parent: dcea8785ad76c477838a2a0cf0bb7b47ea71c356
author: Angie Chiang <angiebird@google.com>
date: Fri Dec 7 17:27:11 EST 2018
Merge changes Id10f72b3,Icde1b01e,I391aa322 * changes: Implement find_prev_nb_full_mvs Implement get_full_mv() Pass mv_num into vp9_nb_mvs_inconsistency()
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -5598,8 +5598,7 @@
&tpl_frame
->tpl_stats_ptr[(mi_row + r) * tpl_frame->stride + mi_col + c];
if (tpl_ptr->ready[rf_idx]) {- nb_full_mvs[i].as_mv.row = tpl_ptr->mv_arr[rf_idx].as_mv.row >> 3;
- nb_full_mvs[i].as_mv.col = tpl_ptr->mv_arr[rf_idx].as_mv.col >> 3;
+ nb_full_mvs[i].as_mv = get_full_mv(&tpl_ptr->mv_arr[rf_idx].as_mv);
} else {nb_full_mvs[i].as_int = INVALID_MV;
}
@@ -5666,7 +5665,7 @@
nb_full_mvs);
vp9_full_pixel_diamond_new(
cpi, x, &best_ref_mv1_full, step_param, lambda, 1, &cpi->fn_ptr[bsize],
- nb_full_mvs, &tpl_stats->mv_arr[rf_idx].as_mv,
+ nb_full_mvs, NB_MVS_NUM, &tpl_stats->mv_arr[rf_idx].as_mv,
&tpl_stats->mv_dist[rf_idx], &tpl_stats->mv_cost[rf_idx]);
#else
(void)frame_idx;
@@ -6361,7 +6360,7 @@
full_mv.row = this_tpl_stats->mv_arr[rf_idx].as_mv.row >> 3;
full_mv.col = this_tpl_stats->mv_arr[rf_idx].as_mv.col >> 3;
this_tpl_stats->mv_cost[rf_idx] =
- av1_nb_mvs_inconsistency(&full_mv, nb_full_mvs);
+ vp9_nb_mvs_inconsistency(&full_mv, nb_full_mvs, NB_MVS_NUM);
#endif // RE_COMPUTE_MV_INCONSISTENCY
tpl_frame->mv_dist_sum[rf_idx] += this_tpl_stats->mv_dist[rf_idx];
tpl_frame->mv_cost_sum[rf_idx] += this_tpl_stats->mv_cost[rf_idx];
--- a/vp9/encoder/vp9_mcomp.c
+++ b/vp9/encoder/vp9_mcomp.c
@@ -1733,12 +1733,13 @@
}
#if CONFIG_NON_GREEDY_MV
-double av1_nb_mvs_inconsistency(const MV *mv, const int_mv *nb_mvs) {+double vp9_nb_mvs_inconsistency(const MV *mv, const int_mv *nb_mvs,
+ int mv_num) {int i;
int update = 0;
double best_cost = 0;
vpx_clear_system_state();
- for (i = 0; i < NB_MVS_NUM; ++i) {+ for (i = 0; i < mv_num; ++i) { if (nb_mvs[i].as_int != INVALID_MV) {MV nb_mv = nb_mvs[i].as_mv;
const double row_diff = mv->row - nb_mv.row;
@@ -1762,7 +1763,7 @@
double *best_mv_dist, double *best_mv_cost,
int search_param, double lambda, int *num00,
const vp9_variance_fn_ptr_t *fn_ptr,
- const int_mv *nb_full_mvs) {+ const int_mv *nb_full_mvs, int full_mv_num) {int i, j, step;
const MACROBLOCKD *const xd = &x->e_mbd;
@@ -1799,7 +1800,8 @@
// Check the starting position
*best_mv_dist = fn_ptr->sdf(what, what_stride, in_what, in_what_stride);
- *best_mv_cost = av1_nb_mvs_inconsistency(best_full_mv, nb_full_mvs);
+ *best_mv_cost =
+ vp9_nb_mvs_inconsistency(best_full_mv, nb_full_mvs, full_mv_num);
bestsad = (*best_mv_dist) + lambda * (*best_mv_cost);
i = 0;
@@ -1833,7 +1835,7 @@
best_full_mv->col + ss_mv[i].col };
const double mv_dist = sad_array[t];
const double mv_cost =
- av1_nb_mvs_inconsistency(&this_mv, nb_full_mvs);
+ vp9_nb_mvs_inconsistency(&this_mv, nb_full_mvs, full_mv_num);
double thissad = mv_dist + lambda * mv_cost;
if (thissad < bestsad) {bestsad = thissad;
@@ -1854,7 +1856,7 @@
const double mv_dist =
fn_ptr->sdf(what, what_stride, check_here, in_what_stride);
const double mv_cost =
- av1_nb_mvs_inconsistency(&this_mv, nb_full_mvs);
+ vp9_nb_mvs_inconsistency(&this_mv, nb_full_mvs, full_mv_num);
double thissad = mv_dist + lambda * mv_cost;
if (thissad < bestsad) {bestsad = thissad;
@@ -2242,16 +2244,17 @@
MV *mvp_full, int step_param, double lambda,
int do_refine,
const vp9_variance_fn_ptr_t *fn_ptr,
- const int_mv *nb_full_mvs, MV *best_mv,
- double *best_mv_dist, double *best_mv_cost) {+ const int_mv *nb_full_mvs, int full_mv_num,
+ MV *best_mv, double *best_mv_dist,
+ double *best_mv_cost) {int n, num00 = 0;
double thissme;
double bestsme;
const int further_steps = MAX_MVSEARCH_STEPS - 1 - step_param;
vpx_clear_system_state();
- bestsme = vp9_diamond_search_sad_new(x, &cpi->ss_cfg, mvp_full, best_mv,
- best_mv_dist, best_mv_cost, step_param,
- lambda, &n, fn_ptr, nb_full_mvs);
+ bestsme = vp9_diamond_search_sad_new(
+ x, &cpi->ss_cfg, mvp_full, best_mv, best_mv_dist, best_mv_cost,
+ step_param, lambda, &n, fn_ptr, nb_full_mvs, full_mv_num);
// If there won't be more n-step search, check to see if refining search is
// needed.
@@ -2265,9 +2268,9 @@
MV temp_mv;
double mv_dist;
double mv_cost;
- thissme = vp9_diamond_search_sad_new(x, &cpi->ss_cfg, mvp_full, &temp_mv,
- &mv_dist, &mv_cost, step_param + n,
- lambda, &num00, fn_ptr, nb_full_mvs);
+ thissme = vp9_diamond_search_sad_new(
+ x, &cpi->ss_cfg, mvp_full, &temp_mv, &mv_dist, &mv_cost,
+ step_param + n, lambda, &num00, fn_ptr, nb_full_mvs, full_mv_num);
// check to see if refining search is needed.
if (num00 > further_steps - n) do_refine = 0;
@@ -2286,9 +2289,9 @@
MV temp_mv = *best_mv;
double mv_dist;
double mv_cost;
- thissme =
- vp9_refining_search_sad_new(x, &temp_mv, &mv_dist, &mv_cost, lambda,
- search_range, fn_ptr, nb_full_mvs);
+ thissme = vp9_refining_search_sad_new(x, &temp_mv, &mv_dist, &mv_cost,
+ lambda, search_range, fn_ptr,
+ nb_full_mvs, full_mv_num);
if (thissme < bestsme) {bestsme = thissme;
*best_mv = temp_mv;
@@ -2428,7 +2431,7 @@
double *best_mv_dist, double *best_mv_cost,
double lambda, int search_range,
const vp9_variance_fn_ptr_t *fn_ptr,
- const int_mv *nb_full_mvs) {+ const int_mv *nb_full_mvs, int full_mv_num) {const MACROBLOCKD *const xd = &x->e_mbd;
const MV neighbors[4] = { { -1, 0 }, { 0, -1 }, { 0, 1 }, { 1, 0 } };const struct buf_2d *const what = &x->plane[0].src;
@@ -2439,7 +2442,8 @@
vpx_clear_system_state();
*best_mv_dist =
fn_ptr->sdf(what->buf, what->stride, best_address, in_what->stride);
- *best_mv_cost = av1_nb_mvs_inconsistency(best_full_mv, nb_full_mvs);
+ *best_mv_cost =
+ vp9_nb_mvs_inconsistency(best_full_mv, nb_full_mvs, full_mv_num);
best_sad = (*best_mv_dist) + lambda * (*best_mv_cost);
for (i = 0; i < search_range; i++) {@@ -2461,7 +2465,8 @@
const MV mv = { best_full_mv->row + neighbors[j].row,best_full_mv->col + neighbors[j].col };
const double mv_dist = sads[j];
- const double mv_cost = av1_nb_mvs_inconsistency(&mv, nb_full_mvs);
+ const double mv_cost =
+ vp9_nb_mvs_inconsistency(&mv, nb_full_mvs, full_mv_num);
const double thissad = mv_dist + lambda * mv_cost;
if (thissad < best_sad) {best_sad = thissad;
@@ -2479,7 +2484,8 @@
const double mv_dist =
fn_ptr->sdf(what->buf, what->stride,
get_buf_from_mv(in_what, &mv), in_what->stride);
- const double mv_cost = av1_nb_mvs_inconsistency(&mv, nb_full_mvs);
+ const double mv_cost =
+ vp9_nb_mvs_inconsistency(&mv, nb_full_mvs, full_mv_num);
const double thissad = mv_dist + lambda * mv_cost;
if (thissad < best_sad) {best_sad = thissad;
--- a/vp9/encoder/vp9_mcomp.h
+++ b/vp9/encoder/vp9_mcomp.h
@@ -126,16 +126,23 @@
double *best_mv_dist, double *best_mv_cost,
double lambda, int search_range,
const vp9_variance_fn_ptr_t *fn_ptr,
- const int_mv *nb_full_mvs);
+ const int_mv *nb_full_mvs, int full_mv_num);
double vp9_full_pixel_diamond_new(const struct VP9_COMP *cpi, MACROBLOCK *x,
MV *mvp_full, int step_param, double lambda,
int do_refine,
const vp9_variance_fn_ptr_t *fn_ptr,
- const int_mv *nb_full_mvs, MV *best_mv,
- double *best_mv_dist, double *best_mv_cost);
+ const int_mv *nb_full_mvs, int full_mv_num,
+ MV *best_mv, double *best_mv_dist,
+ double *best_mv_cost);
-double av1_nb_mvs_inconsistency(const MV *mv, const int_mv *nb_mvs);
+double vp9_nb_mvs_inconsistency(const MV *mv, const int_mv *nb_mvs, int mv_num);
+static INLINE MV get_full_mv(const MV *mv) {+ MV out_mv;
+ out_mv.row = mv->row >> 3;
+ out_mv.col = mv->col >> 3;
+ return out_mv;
+}
#endif // CONFIG_NON_GREEDY_MV
#ifdef __cplusplus
} // extern "C"
--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -2315,6 +2315,61 @@
block_size);
}
+#if CONFIG_NON_GREEDY_MV
+#define MAX_PREV_NB_FULL_MV_NUM 8
+static int find_prev_nb_full_mvs(const VP9_COMMON *cm, const MACROBLOCKD *xd,
+ int ref_frame, BLOCK_SIZE bsize, int mi_row,
+ int mi_col, int_mv *nb_full_mvs) {+ int i;
+ const TileInfo *tile = &xd->tile;
+ int full_mv_num = 0;
+ assert(bsize >= BLOCK_8X8);
+ for (i = 0; i < MVREF_NEIGHBOURS; ++i) {+ const POSITION *mv_ref = &mv_ref_blocks[bsize][i];
+ if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {+ const MODE_INFO *nb_mi =
+ xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
+ if (nb_mi->sb_type >= BLOCK_8X8) {+ if (nb_mi->ref_frame[0] == ref_frame) {+ nb_full_mvs[full_mv_num].as_mv = get_full_mv(&nb_mi->mv[0].as_mv);
+ ++full_mv_num;
+ if (full_mv_num >= MAX_PREV_NB_FULL_MV_NUM) {+ return full_mv_num;
+ }
+ } else if (nb_mi->ref_frame[1] == ref_frame) {+ nb_full_mvs[full_mv_num].as_mv = get_full_mv(&nb_mi->mv[1].as_mv);
+ ++full_mv_num;
+ if (full_mv_num >= MAX_PREV_NB_FULL_MV_NUM) {+ return full_mv_num;
+ }
+ }
+ } else {+ int j;
+ for (j = 0; j < 4; ++j) {+ // TODO(angiebird): avoid using duplicated mvs
+ if (nb_mi->ref_frame[0] == ref_frame) {+ nb_full_mvs[full_mv_num].as_mv =
+ get_full_mv(&nb_mi->bmi[j].as_mv[0].as_mv);
+ ++full_mv_num;
+ if (full_mv_num >= MAX_PREV_NB_FULL_MV_NUM) {+ return full_mv_num;
+ }
+ } else if (nb_mi->ref_frame[1] == ref_frame) {+ nb_full_mvs[full_mv_num].as_mv =
+ get_full_mv(&nb_mi->bmi[j].as_mv[1].as_mv);
+ ++full_mv_num;
+ if (full_mv_num >= MAX_PREV_NB_FULL_MV_NUM) {+ return full_mv_num;
+ }
+ }
+ }
+ }
+ }
+ }
+ return full_mv_num;
+}
+#endif // CONFIG_NON_GREEDY_MV
+
static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
int mi_row, int mi_col, int_mv *tmp_mv,
int *rate_mv) {@@ -2338,11 +2393,12 @@
#if CONFIG_NON_GREEDY_MV
double mv_dist = 0;
double mv_cost = 0;
- double lambda = 0;
+ double lambda = (pw * ph) / 4;
double bestsme;
- int_mv nb_full_mvs[NB_MVS_NUM];
- // TODO(angiebird): Set nb_full_mvs properly.
- vp9_zero(nb_full_mvs);
+ int_mv nb_full_mvs[MAX_PREV_NB_FULL_MV_NUM];
+
+ const int nb_full_mv_num =
+ find_prev_nb_full_mvs(cm, xd, ref, bsize, mi_row, mi_col, nb_full_mvs);
#else // CONFIG_NON_GREEDY_MV
int bestsme = INT_MAX;
int sadpb = x->sadperbit16;
@@ -2418,9 +2474,9 @@
mvp_full.row >>= 3;
#if CONFIG_NON_GREEDY_MV
- bestsme = vp9_full_pixel_diamond_new(cpi, x, &mvp_full, step_param, lambda, 1,
- &cpi->fn_ptr[bsize], nb_full_mvs,
- &tmp_mv->as_mv, &mv_dist, &mv_cost);
+ bestsme = vp9_full_pixel_diamond_new(
+ cpi, x, &mvp_full, step_param, lambda, 1, &cpi->fn_ptr[bsize],
+ nb_full_mvs, nb_full_mv_num, &tmp_mv->as_mv, &mv_dist, &mv_cost);
#else // CONFIG_NON_GREEDY_MV
bestsme = vp9_full_pixel_search(
cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method, sadpb,
@@ -2461,8 +2517,8 @@
#if CONFIG_NON_GREEDY_MV
this_me = vp9_full_pixel_diamond_new(
cpi, x, &mvp_full, VPXMAX(step_param, MAX_MVSEARCH_STEPS - step),
- lambda, 1, &cpi->fn_ptr[bsize], nb_full_mvs, &this_mv, &mv_dist,
- &mv_cost);
+ lambda, 1, &cpi->fn_ptr[bsize], nb_full_mvs, nb_full_mv_num, &this_mv,
+ &mv_dist, &mv_cost);
#else // CONFIG_NON_GREEDY_MV
this_me = vp9_full_pixel_search(
cpi, x, bsize, &mvp_full,