hashirama/uvg266
1
0
Fork 0
mirror of https://github.com/ultravideo/uvg266.git synced 2024-11-24 10:34:05 +00:00
Code Issues Actions Projects Releases Wiki Activity

Compare rough cost of the best merge mode against AMVP to make mode decision

Browse source
This commit is contained in:
Ari Lemmetti 2019-08-26 21:01:25 +03:00
parent 8f866ff83a
commit aa8ab195d1
1 changed files with 69 additions and 61 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

128
src/search_inter.c
View file

@ -1460,22 +1460,13 @@ static bool merge_candidate_in_list(inter_merge_cand_t * all_cands,
for (int i = 0; i < list_size && !found; ++i) { for (int i = 0; i < list_size && !found; ++i) {
inter_merge_cand_t * list_cand = &all_cands[added_idx_list[i]]; inter_merge_cand_t * list_cand = &all_cands[added_idx_list[i]];
// Check only the specified dir(s) in case found = cand_to_add->dir == list_cand->dir &&
// something unused is left uninitialized
found = cand_to_add->dir == list_cand->dir;
if (cand_to_add->dir & 1) {
found = found &&
cand_to_add->ref[0] == list_cand->ref[0] && cand_to_add->ref[0] == list_cand->ref[0] &&
cand_to_add->mv[0][0] == list_cand->mv[0][0] && cand_to_add->mv[0][0] == list_cand->mv[0][0] &&
cand_to_add->mv[0][1] == list_cand->mv[0][1]; cand_to_add->mv[0][1] == list_cand->mv[0][1] &&
}
if (cand_to_add->dir & 2) {
found = found &&
cand_to_add->ref[1] == list_cand->ref[1] && cand_to_add->ref[1] == list_cand->ref[1] &&
cand_to_add->mv[1][0] == list_cand->mv[1][0] && cand_to_add->mv[1][0] == list_cand->mv[1][0] &&
cand_to_add->mv[1][1] == list_cand->mv[1][1]; cand_to_add->mv[1][1] == list_cand->mv[1][1];
}
} }
return found; return found;
@ -1550,16 +1541,17 @@ static void search_pu_inter(encoder_state_t * const state,
CU_SET_MV_CAND(cur_cu, 0, 0); CU_SET_MV_CAND(cur_cu, 0, 0);
CU_SET_MV_CAND(cur_cu, 1, 0); CU_SET_MV_CAND(cur_cu, 1, 0);
// Early Skip Mode Decision // Merge Analysis starts here
if (cfg->early_skip && cur_cu->part_size == SIZE_2Nx2N) { int8_t mrg_cands[MRG_MAX_NUM_CANDS];
double mrg_costs[MRG_MAX_NUM_CANDS];
for (int i = 0; i < MRG_MAX_NUM_CANDS; ++i) {
mrg_cands[i] = -1;
mrg_costs[i] = MAX_DOUBLE;
}
if (cur_cu->part_size == SIZE_2Nx2N) {
int num_rdo_cands = 0; int num_rdo_cands = 0;
int8_t mrg_cands[MRG_MAX_NUM_CANDS];
double mrg_costs[MRG_MAX_NUM_CANDS];
for (int i = 0; i < MRG_MAX_NUM_CANDS; ++i) {
mrg_cands[i] = -1;
mrg_costs[i] = MAX_DOUBLE;
}
// Check motion vector constraints and perform rough search // Check motion vector constraints and perform rough search
for (int merge_idx = 0; merge_idx < info.num_merge_cand; ++merge_idx) { for (int merge_idx = 0; merge_idx < info.num_merge_cand; ++merge_idx) {
@ -1573,6 +1565,7 @@ static void search_pu_inter(encoder_state_t * const state,
cur_cu->inter.mv[1][0] = cur_cand->mv[1][0]; cur_cu->inter.mv[1][0] = cur_cand->mv[1][0];
cur_cu->inter.mv[1][1] = cur_cand->mv[1][1]; cur_cu->inter.mv[1][1] = cur_cand->mv[1][1];
if (cur_cu->inter.mv_dir == 3 && !state->encoder_control->cfg.bipred) continue;
bool is_duplicate = merge_candidate_in_list(info.merge_cand, cur_cand, bool is_duplicate = merge_candidate_in_list(info.merge_cand, cur_cand,
mrg_cands, mrg_cands,
num_rdo_cands); num_rdo_cands);
@ -1586,65 +1579,62 @@ static void search_pu_inter(encoder_state_t * const state,
continue; continue;
} }
if (cfg->rdo >= 2) { kvz_inter_recon_cu(state, lcu, x, y, width);
mrg_costs[num_rdo_cands] = kvz_satd_any_size(width, height,
kvz_inter_recon_cu(state, lcu, x, y, width); lcu->rec.y + y_local * LCU_WIDTH + x_local, LCU_WIDTH,
mrg_costs[num_rdo_cands] = kvz_satd_any_size(width, height, lcu->ref.y + y_local * LCU_WIDTH + x_local, LCU_WIDTH);
lcu->rec.y + y_local * LCU_WIDTH + x_local, LCU_WIDTH,
lcu->ref.y + y_local * LCU_WIDTH + x_local, LCU_WIDTH);
}
mrg_cands[num_rdo_cands] = merge_idx; mrg_cands[num_rdo_cands] = merge_idx;
num_rdo_cands++; num_rdo_cands++;
} }
// Sort candidates by cost
if (cfg->rdo >= 2) { kvz_sort_modes(mrg_cands, mrg_costs, num_rdo_cands);
// Sort candidates by cost
kvz_sort_modes(mrg_cands, mrg_costs, num_rdo_cands);
}
// Limit by availability // Limit by availability
// TODO: Do not limit to just 1 // TODO: Do not limit to just 1
num_rdo_cands = MIN(1, num_rdo_cands); num_rdo_cands = MIN(1, num_rdo_cands);
// RDO search // Early Skip Mode Decision
for (int merge_rdo_idx = 0; merge_rdo_idx < num_rdo_cands; ++merge_rdo_idx) { if (cfg->early_skip) {
for (int merge_rdo_idx = 0; merge_rdo_idx < num_rdo_cands; ++merge_rdo_idx) {
// Reconstruct blocks with merge candidate. // Reconstruct blocks with merge candidate.
// Check luma CBF. Then, check chroma CBFs if luma CBF is not set // Check luma CBF. Then, check chroma CBFs if luma CBF is not set
// and chroma exists. // and chroma exists.
// Early terminate if merge candidate with zero CBF is found. // Early terminate if merge candidate with zero CBF is found.
int merge_idx = mrg_cands[merge_rdo_idx]; int merge_idx = mrg_cands[merge_rdo_idx];
cur_cu->inter.mv_dir = info.merge_cand[merge_idx].dir; cur_cu->inter.mv_dir = info.merge_cand[merge_idx].dir;
cur_cu->inter.mv_ref[0] = info.merge_cand[merge_idx].ref[0]; cur_cu->inter.mv_ref[0] = info.merge_cand[merge_idx].ref[0];
cur_cu->inter.mv_ref[1] = info.merge_cand[merge_idx].ref[1]; cur_cu->inter.mv_ref[1] = info.merge_cand[merge_idx].ref[1];
cur_cu->inter.mv[0][0] = info.merge_cand[merge_idx].mv[0][0]; cur_cu->inter.mv[0][0] = info.merge_cand[merge_idx].mv[0][0];
cur_cu->inter.mv[0][1] = info.merge_cand[merge_idx].mv[0][1]; cur_cu->inter.mv[0][1] = info.merge_cand[merge_idx].mv[0][1];
cur_cu->inter.mv[1][0] = info.merge_cand[merge_idx].mv[1][0]; cur_cu->inter.mv[1][0] = info.merge_cand[merge_idx].mv[1][0];
cur_cu->inter.mv[1][1] = info.merge_cand[merge_idx].mv[1][1]; cur_cu->inter.mv[1][1] = info.merge_cand[merge_idx].mv[1][1];
kvz_lcu_fill_trdepth(lcu, x, y, depth, MAX(1, depth)); kvz_lcu_fill_trdepth(lcu, x, y, depth, MAX(1, depth));
kvz_inter_recon_cu(state, lcu, x, y, width); kvz_inter_recon_cu(state, lcu, x, y, width);
kvz_quantize_lcu_residual(state, true, false, x, y, depth, cur_cu, lcu); kvz_quantize_lcu_residual(state, true, false, x, y, depth, cur_cu, lcu);
if (cbf_is_set(cur_cu->cbf, depth, COLOR_Y)) { if (cbf_is_set(cur_cu->cbf, depth, COLOR_Y)) {
continue; continue;
} }
else if(state->encoder_control->chroma_format != KVZ_CSP_400) { else if (state->encoder_control->chroma_format != KVZ_CSP_400) {
kvz_quantize_lcu_residual(state, false, true, x, y, depth, cur_cu, lcu); kvz_quantize_lcu_residual(state, false, true, x, y, depth, cur_cu, lcu);
if (!cbf_is_set_any(cur_cu->cbf, depth)) { if (!cbf_is_set_any(cur_cu->cbf, depth)) {
cur_cu->type = CU_INTER; cur_cu->type = CU_INTER;
cur_cu->merge_idx = merge_idx; cur_cu->merge_idx = merge_idx;
cur_cu->skipped = true; cur_cu->skipped = true;
*inter_cost = 0.0; // TODO: Check this *inter_cost = 0.0; // TODO: Check this
*inter_bitcost = 0; // TODO: Check this *inter_bitcost = 0; // TODO: Check this
return; return;
}
} }
} }
} }
} }
// AMVP search starts here
for (int ref_idx = 0; ref_idx < state->frame->ref->used_size; ref_idx++) { for (int ref_idx = 0; ref_idx < state->frame->ref->used_size; ref_idx++) {
info.ref_idx = ref_idx; info.ref_idx = ref_idx;
info.ref = state->frame->ref->images[ref_idx]; info.ref = state->frame->ref->images[ref_idx];
@ -1661,6 +1651,24 @@ static void search_pu_inter(encoder_state_t * const state,
search_pu_inter_bipred(&info, depth, lcu, cur_cu, inter_cost, inter_bitcost); search_pu_inter_bipred(&info, depth, lcu, cur_cu, inter_cost, inter_bitcost);
} }
// Compare best merge cost to amvp cost
if (mrg_costs[0] < *inter_cost) {
*inter_cost = mrg_costs[0];
*inter_bitcost = 0; // TODO: Check this
int merge_idx = mrg_cands[0];
cur_cu->type = CU_INTER;
cur_cu->merge_idx = merge_idx;
cur_cu->inter.mv_dir = info.merge_cand[merge_idx].dir;
cur_cu->inter.mv_ref[0] = info.merge_cand[merge_idx].ref[0];
cur_cu->inter.mv_ref[1] = info.merge_cand[merge_idx].ref[1];
cur_cu->inter.mv[0][0] = info.merge_cand[merge_idx].mv[0][0];
cur_cu->inter.mv[0][1] = info.merge_cand[merge_idx].mv[0][1];
cur_cu->inter.mv[1][0] = info.merge_cand[merge_idx].mv[1][0];
cur_cu->inter.mv[1][1] = info.merge_cand[merge_idx].mv[1][1];
cur_cu->merged = true;
cur_cu->skipped = false;
}
if (*inter_cost < INT_MAX && cur_cu->inter.mv_dir == 1) { if (*inter_cost < INT_MAX && cur_cu->inter.mv_dir == 1) {
assert(fracmv_within_tile(&info, cur_cu->inter.mv[0][0], cur_cu->inter.mv[0][1])); assert(fracmv_within_tile(&info, cur_cu->inter.mv[0][0], cur_cu->inter.mv[0][1]));
} }