From 9e644fafd03c073456b24f6ec12376fcb6648ece Mon Sep 17 00:00:00 2001 From: Joose Sainio Date: Wed, 30 Nov 2022 10:43:12 +0200 Subject: [PATCH] [mtt] search with mtt depth 2 and dual tree works without lfnst --- src/cfg.c | 2 +- src/cu.c | 6 +++--- src/cu.h | 2 +- src/encode_coding_tree.c | 20 +++++++++++++------- src/intra.c | 22 +++++++++++++++++++++- src/intra.h | 2 ++ src/search.c | 8 ++++---- src/search_intra.c | 4 ++-- 8 files changed, 47 insertions(+), 19 deletions(-) diff --git a/src/cfg.c b/src/cfg.c index 5ff05859..1bd6587d 100644 --- a/src/cfg.c +++ b/src/cfg.c @@ -227,7 +227,7 @@ int uvg_config_init(uvg_config *cfg) cfg->min_qt_size[1] = 4; cfg->min_qt_size[2] = 4; - cfg->max_btt_depth[0] = 1; + cfg->max_btt_depth[0] = 2; cfg->max_btt_depth[1] = 0; cfg->max_btt_depth[2] = 1; diff --git a/src/cu.c b/src/cu.c index 1e349bc7..0c8dd9f7 100644 --- a/src/cu.c +++ b/src/cu.c @@ -400,7 +400,7 @@ int uvg_get_possible_splits(const encoder_state_t * const state, bool can_btt = split_tree.mtt_depth < max_btd; const enum split_type last_split = (split_tree.split_tree >> (split_tree.current_depth * 3 - 3)) & 7; - const enum split_type parl_split = last_split == BT_HOR_SPLIT ? BT_HOR_SPLIT : BT_VER_SPLIT; + const enum split_type parl_split = last_split == TT_HOR_SPLIT ? BT_HOR_SPLIT : BT_VER_SPLIT; // don't allow QT-splitting below a BT split if (split_tree.current_depth != 0 && last_split != QT_SPLIT /* && !(width > 64 || height > 64)*/) splits[QT_SPLIT] = false; @@ -459,12 +459,12 @@ int uvg_get_possible_splits(const encoder_state_t * const state, //if (modeType == MODE_TYPE_INTER && width * height == 32) splits[BT_VER_SPLIT] = splits[BT_HOR_SPLIT] = false; - if (height <= 2 * min_tt_size || height > max_tt_size || width > max_tt_size) + if (cu_loc->chroma_height <= min_tt_size || height > max_tt_size || width > max_tt_size) splits[TT_HOR_SPLIT] = false; if (width > 64 || height > 64) splits[TT_HOR_SPLIT] = false; if (tree_type == UVG_CHROMA_T && width * height <= 16 * 2) splits[TT_HOR_SPLIT] = false; - if (width <= 2 * min_tt_size || width > max_tt_size || height > max_tt_size) + if (cu_loc->chroma_width <= min_tt_size || width > max_tt_size || height > max_tt_size) splits[TT_VER_SPLIT] = false; if (width > 64 || height > 64) splits[TT_VER_SPLIT] = false; if (tree_type == UVG_CHROMA_T && (width * height <= 16 * 2 || width == 8)) splits[TT_VER_SPLIT] = false; diff --git a/src/cu.h b/src/cu.h index bfbb50a1..7722852d 100644 --- a/src/cu.h +++ b/src/cu.h @@ -607,7 +607,7 @@ static INLINE void cbf_copy(uint16_t *cbf, uint16_t src, color_t plane) *cbf |= src & (1 << plane); } -#define GET_SPLITDATA(CU,curDepth) (((CU)->split_tree >> (curDepth)) & 7) +#define GET_SPLITDATA(CU,curDepth) (((CU)->split_tree >> ((curDepth) * 3)) & 7) #define PU_IS_TU(cu) ((cu)->log2_width <= TR_MAX_LOG2_SIZE && (cu)->log2_height <= TR_MAX_LOG2_SIZE) #endif diff --git a/src/encode_coding_tree.c b/src/encode_coding_tree.c index ed2e8468..b73a446c 100644 --- a/src/encode_coding_tree.c +++ b/src/encode_coding_tree.c @@ -563,6 +563,7 @@ static void encode_transform_unit( (cu_info_t * )cur_pu, NULL); } + if (tree_type == UVG_LUMA_T) return; } bool joint_chroma = cur_pu->joint_cb_cr != 0; @@ -627,8 +628,9 @@ static void encode_transform_coeff( cur_tu = uvg_cu_array_at_const(used_array, x, y); } - const bool ver_split = cu_loc->height > TR_MAX_WIDTH; - const bool hor_split = cu_loc->width > TR_MAX_WIDTH; + const int tr_limit = (TR_MAX_WIDTH >> (tree_type == UVG_CHROMA_T)); + const bool ver_split = cu_loc->height > tr_limit; + const bool hor_split = cu_loc->width > tr_limit; const int cb_flag_y = tree_type != UVG_CHROMA_T ? cbf_is_set(cur_tu->cbf, COLOR_Y) : 0; const int cb_flag_u = tree_type != UVG_LUMA_T ?(cur_tu->joint_cb_cr ? (cur_tu->joint_cb_cr >> 1) & 1 : cbf_is_set(cur_tu->cbf, COLOR_U)) : 0; @@ -637,10 +639,10 @@ static void encode_transform_coeff( if (hor_split || ver_split) { enum split_type split; - if (cu_loc->width > TR_MAX_WIDTH && cu_loc->height > TR_MAX_WIDTH) { + if (cu_loc->width > tr_limit && cu_loc->height > tr_limit) { split = QT_SPLIT; } - else if (cu_loc->width > TR_MAX_WIDTH) { + else if (cu_loc->width > tr_limit) { split = BT_VER_SPLIT; } else { @@ -650,6 +652,10 @@ static void encode_transform_coeff( cu_loc_t split_cu_loc[4]; const int split_count = uvg_get_split_locs(cu_loc, split, split_cu_loc,NULL); for (int i = 0; i < split_count; ++i) { + if(tree_type == UVG_CHROMA_T) { + split_cu_loc[i].chroma_width = split_cu_loc[i].width; + split_cu_loc[i].chroma_height = split_cu_loc[i].height; + } encode_transform_coeff(state, &split_cu_loc[i], only_chroma, coeff, NULL, tree_type, true, false, luma_cbf_ctx, &split_cu_loc[i], &split_cu_loc[i]); } @@ -1252,7 +1258,7 @@ uint8_t uvg_write_split_flag( } - if (!is_implicit && can_split[QT_SPLIT] && (can_split[BT_HOR_SPLIT] || can_split[BT_VER_SPLIT] || can_split[TT_HOR_SPLIT] || can_split[TT_VER_SPLIT]) && split_flag != NO_SPLIT) { + if (!is_implicit && (can_split[BT_HOR_SPLIT] || can_split[BT_VER_SPLIT] || can_split[TT_HOR_SPLIT] || can_split[TT_VER_SPLIT]) && split_flag != NO_SPLIT) { bool qt_split = split_flag == QT_SPLIT; if((can_split[BT_VER_SPLIT] || can_split[BT_HOR_SPLIT] || can_split[TT_VER_SPLIT] || can_split[TT_HOR_SPLIT]) && can_split[QT_SPLIT]) { unsigned left_qt_depth = 0; @@ -1617,7 +1623,7 @@ void uvg_encode_coding_tree( has_chroma) || tree_type == UVG_CHROMA_T) && tree_type != UVG_LUMA_T) { int8_t luma_dir = uvg_get_co_located_luma_mode(tree_type != UVG_CHROMA_T ? chroma_loc : cu_loc, cu_loc, cur_cu, NULL, frame->cu_array, UVG_CHROMA_T); - encode_chroma_intra_cu(cabac, cur_cu, state->encoder_control->cfg.cclm, luma_dir,NULL); + encode_chroma_intra_cu(cabac, cur_cu, state->encoder_control->cfg.cclm && uvg_cclm_is_allowed(state, cu_loc, cur_cu, tree_type), luma_dir,NULL); // LFNST constraints must be reset here. Otherwise the left over values will interfere when calculating new constraints cu_info_t* tmp = (cu_info_t*)cur_cu; tmp->violates_lfnst_constrained_luma = false; @@ -1769,7 +1775,7 @@ double uvg_mock_encode_coding_unit( int8_t luma_dir = uvg_get_co_located_luma_mode(chroma_loc,cu_loc , cur_cu, tree_type != UVG_CHROMA_T ? lcu : NULL, tree_type == UVG_CHROMA_T ? state->tile->frame->cu_array : NULL, is_separate_tree ? UVG_CHROMA_T : tree_type); - encode_chroma_intra_cu(cabac, cur_cu, state->encoder_control->cfg.cclm, luma_dir, &bits); + encode_chroma_intra_cu(cabac, cur_cu, state->encoder_control->cfg.cclm && uvg_cclm_is_allowed(state, cu_loc, cur_cu, tree_type), luma_dir, &bits); } } else { diff --git a/src/intra.c b/src/intra.c index 1dd74cb3..09c66c13 100644 --- a/src/intra.c +++ b/src/intra.c @@ -283,6 +283,26 @@ static void intra_pred_dc( } +bool uvg_cclm_is_allowed(const encoder_state_t* const state, const cu_loc_t * const luma_loc, cu_info_t const * const cur_cu, enum + uvg_tree_type tree_type) +{ + if (tree_type != UVG_CHROMA_T) { + return true; + } + uint32_t chroma_split = GET_SPLITDATA(cur_cu, 0); + if((chroma_split == BT_VER_SPLIT || chroma_split == TT_VER_SPLIT || chroma_split == TT_HOR_SPLIT) && GET_SPLITDATA(cur_cu, 1) == NO_SPLIT) return false; + const cu_info_t* const luma_cu = uvg_cu_array_at_const(state->tile->frame->cu_array, luma_loc->x, luma_loc->y); + uint32_t split = GET_SPLITDATA(luma_cu, 0); + if (split != QT_SPLIT && split != NO_SPLIT) { + return false; + } + if (split != NO_SPLIT && luma_cu->intra.isp_mode != ISP_MODE_NO_ISP) { + return false; + } + return true; +} + + enum lm_mode { LM_CHROMA_IDX = 81, @@ -1846,7 +1866,7 @@ void uvg_intra_recon_cu( const uint8_t depth = 6 - uvg_g_convert_to_log2[cu_loc->width]; const vector2d_t lcu_px = { cu_loc->local_x >> (tree_type == UVG_CHROMA_T), cu_loc->local_y >> (tree_type == UVG_CHROMA_T) }; const int8_t width = cu_loc->width; - const int8_t height = cu_loc->height; // TODO: height for non-square blocks. + const int8_t height = cu_loc->height; if (cur_cu == NULL) { cur_cu = LCU_GET_CU_AT_PX(lcu, lcu_px.x, lcu_px.y); } diff --git a/src/intra.h b/src/intra.h index fc81e645..5d7d84e7 100644 --- a/src/intra.h +++ b/src/intra.h @@ -158,6 +158,8 @@ int8_t uvg_get_co_located_luma_mode( const lcu_t* const lcu, const cu_array_t* const cu_array, enum uvg_tree_type tree_type); +bool uvg_cclm_is_allowed(const encoder_state_t* const state, const cu_loc_t* const luma_loc, cu_info_t const* const cur_cu, enum + uvg_tree_type tree_type); uint8_t uvg_get_mip_flag_context( const cu_loc_t* const cu_loc, diff --git a/src/search.c b/src/search.c index 3c76dc93..9546cc0a 100644 --- a/src/search.c +++ b/src/search.c @@ -215,7 +215,7 @@ static void work_tree_copy_up( copy_cu_info (from, to, cu_loc, tree_type); copy_cu_pixels(from, to, cu_loc, cu_loc != chroma_loc && tree_type == UVG_LUMA_T ? UVG_LUMA_T : tree_type); copy_cu_coeffs(cu_loc, from, to, joint, cu_loc != chroma_loc && tree_type == UVG_LUMA_T ? UVG_LUMA_T : tree_type); - if (cu_loc != chroma_loc && tree_type == UVG_LUMA_T) { + if (cu_loc != chroma_loc && tree_type != UVG_LUMA_T) { copy_cu_pixels(from, to, chroma_loc, UVG_CHROMA_T); copy_cu_coeffs(chroma_loc, from, to, joint, UVG_CHROMA_T); } @@ -1189,7 +1189,7 @@ static double search_cu( bool recon_chroma = true; bool recon_luma = tree_type != UVG_CHROMA_T; - if (is_separate_tree || !has_chroma || state->encoder_control->chroma_format == UVG_CSP_400 || tree_type == UVG_LUMA_T) { + if (is_separate_tree || !has_chroma || state->encoder_control->chroma_format == UVG_CSP_400 || tree_type == UVG_LUMA_T || cu_loc->chroma_height % 4 == 2) { recon_chroma = false; } lcu_fill_cu_info(lcu, x_local, y_local, cu_width, cu_height, cur_cu); @@ -1200,7 +1200,7 @@ static double search_cu( recon_luma, recon_chroma); - if((is_separate_tree && has_chroma && tree_type != UVG_LUMA_T && state->encoder_control->chroma_format != UVG_CSP_400 ) + if((!recon_chroma && state->encoder_control->chroma_format != UVG_CSP_400 ) || tree_type == UVG_CHROMA_T) { intra_search.pred_cu.intra.mode_chroma = cur_cu->intra.mode_chroma; uvg_intra_recon_cu(state, @@ -1360,7 +1360,7 @@ static double search_cu( cabac_data_t best_split_cabac; memcpy(&post_seach_cabac, &state->search_cabac, sizeof(post_seach_cabac)); for (int split_type = QT_SPLIT; split_type <= TT_VER_SPLIT; ++split_type) { - if (!can_split[split_type] || (split_type != QT_SPLIT && depth == 0) || (split_type == QT_SPLIT && depth == 1)) continue; + if (!can_split[split_type] || (tree_type == UVG_CHROMA_T && split_type == TT_HOR_SPLIT && cu_loc->chroma_height == 8)) continue; split_tree_t new_split = { split_tree.split_tree | split_type << (split_tree.current_depth * 3), split_tree.current_depth + 1, diff --git a/src/search_intra.c b/src/search_intra.c index 2a406076..4ee36f95 100644 --- a/src/search_intra.c +++ b/src/search_intra.c @@ -1476,7 +1476,7 @@ int8_t uvg_search_intra_chroma_rdo( } pred_cu->cr_lfnst_idx = lfnst; chroma_data[mode_i].lfnst_costs[lfnst] += mode_bits * state->lambda; - if (PU_IS_TU(pred_cu)) { + if (PU_IS_TU(pred_cu) && (tree_type != UVG_CHROMA_T || (pred_cu->log2_width < 5 && pred_cu->log2_height < 5))) { uvg_intra_predict( state, &refs[COLOR_U - 1], @@ -1594,7 +1594,7 @@ int8_t uvg_search_cu_intra_chroma( const cu_info_t *cur_pu = &search_data->pred_cu; int8_t modes[8] = { 0, 50, 18, 1, luma_mode, 81, 82, 83 }; - uint8_t total_modes = (state->encoder_control->cfg.cclm ? 8 : 5); + uint8_t total_modes = (state->encoder_control->cfg.cclm && uvg_cclm_is_allowed(state, cu_loc, cur_pu, tree_type) ? 8 : 5); for(int i = 0; i < 4; i++) { if (modes[i] == luma_mode) { modes[i] = 66;