[mtt] search with depth 1 mtt kinda working

src/cu.c

@@ -399,11 +399,11 @@ int uvg_get_possible_splits(const encoder_state_t * const state,
   splits[NO_SPLIT] = splits[QT_SPLIT] = splits[BT_HOR_SPLIT] = splits[TT_HOR_SPLIT] = splits[BT_VER_SPLIT] = splits[TT_VER_SPLIT] = true;
   bool can_btt = split_tree.mtt_depth < max_btd;
   
-  const enum split_type last_split = (split_tree.split_tree >> (split_tree.current_depth * 3)) & 7;
+  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;
 
   // 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;
+  if (split_tree.current_depth != 0 && last_split != QT_SPLIT /* && !(width > 64 || height > 64)*/) splits[QT_SPLIT] = false;
   if (width <= min_qt_size)                              splits[QT_SPLIT] = false;
 
   if (tree_type == UVG_CHROMA_T && width <= 4) splits[QT_SPLIT] = false;
@@ -488,6 +488,7 @@ int uvg_count_available_edge_cus(const cu_loc_t* const cu_loc, const lcu_t* cons
 
   int amount = 0;
   if(left) {
+    if (cu_loc->local_y == 0 && cu_loc->local_x == 32 && cu_loc->height == 32 && cu_loc->width == 32) return 8;
     while (LCU_GET_CU_AT_PX(lcu, cu_loc->local_x - TR_MIN_WIDTH, cu_loc->local_y + amount)->type != CU_NOTSET && (cu_loc->local_y + amount) < LCU_WIDTH) {
       amount += TR_MIN_WIDTH;
     }

src/encode_coding_tree.c

@@ -1643,7 +1643,7 @@ void uvg_encode_coding_tree(
       ((is_local_dual_tree &&
       has_chroma) || tree_type == UVG_CHROMA_T) &&
       tree_type != UVG_LUMA_T)   {
-      int8_t luma_dir = uvg_get_co_located_luma_mode(chroma_loc->x, chroma_loc->y, chroma_loc->width, chroma_loc->height, NULL, frame->cu_array, UVG_CHROMA_T);
+      int8_t luma_dir = uvg_get_co_located_luma_mode(chroma_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);
       // 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;
@@ -1793,9 +1793,9 @@ double uvg_mock_encode_coding_unit(
       uvg_encode_intra_luma_coding_unit(state, cabac, cur_cu, cu_loc, lcu, &bits);
     }
     if((chroma_loc || tree_type == UVG_CHROMA_T) && state->encoder_control->chroma_format != UVG_CSP_400 && tree_type != UVG_LUMA_T) {
-      int8_t luma_dir = uvg_get_co_located_luma_mode(chroma_loc->x, chroma_loc->y, chroma_loc->width, chroma_loc->height,
+      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 ? lcu : NULL,
+              tree_type == UVG_CHROMA_T ? state->tile->frame->cu_array : NULL,
-        tree_type == UVG_CHROMA_T ? state->tile->frame->cu_array : NULL,   is_separate_tree ? UVG_CHROMA_T : tree_type);
+              is_separate_tree ? UVG_CHROMA_T : tree_type);
       encode_chroma_intra_cu(cabac, cur_cu, state->encoder_control->cfg.cclm, luma_dir, &bits);
     }
   }

src/intra.c

@@ -1603,22 +1603,26 @@ void uvg_intra_predict(
 
 // This function works on luma coordinates 
 int8_t uvg_get_co_located_luma_mode(
-  int x,
+  const cu_loc_t* const chroma_loc,
-  int y,
+  const cu_loc_t* const cu_loc,
-  int width,
+  const cu_info_t* luma_cu,
-  int height,
   const lcu_t* const lcu,
   const cu_array_t* const cu_array,
   enum uvg_tree_type tree_type)
 {
+  int x = chroma_loc->x;
+  int y = chroma_loc->y;
   assert((cu_array || lcu) && !(cu_array && lcu));
   assert(tree_type != UVG_LUMA_T && "Luma only CU shouldn't need colocated luma CU");
   if(tree_type == UVG_CHROMA_T) {
-    x += width >> 1;
+    x += chroma_loc->width >> 1;
-    y += height >> 1;
+    y += chroma_loc->height >> 1;
   }
   const cu_info_t* cu;
-  if(cu_array) {
+  if (lcu && cu_loc->x <= x && x < cu_loc->x + cu_loc->width && cu_loc->y <= y && y < cu_loc->y + cu_loc->height) {
+    cu = luma_cu;
+  }
+  else if(cu_array) {
     cu = uvg_cu_array_at_const(cu_array, x, y);
   }
   else {

src/intra.h

@@ -152,10 +152,9 @@ void uvg_intra_recon_cu(
   bool recon_chroma);
 
 int8_t uvg_get_co_located_luma_mode(
-  int x,
+  const cu_loc_t* const chroma_loc,
-  int y,
+  const cu_loc_t* const cu_loc,
-  int width,
+  const cu_info_t* luma_cu,
-  int height,
   const lcu_t* const lcu,
   const cu_array_t* const cu_array,
   enum uvg_tree_type tree_type);

src/search.c

@@ -1107,9 +1107,10 @@ static double search_cu(
 
           intra_search.pred_cu.joint_cb_cr = 0;
           if(tree_type == UVG_CHROMA_T || is_separate_tree) {
-            intra_mode = uvg_get_co_located_luma_mode(chroma_loc->x, chroma_loc->y, chroma_loc->width, chroma_loc->height,
+            intra_mode = uvg_get_co_located_luma_mode(
-                                                    is_separate_tree ? lcu : NULL,
+                    chroma_loc, cu_loc, &intra_search.pred_cu, is_separate_tree ? lcu : NULL,
-                                                    tree_type == UVG_CHROMA_T ? state->tile->frame->cu_array : NULL, UVG_CHROMA_T);
+                    tree_type == UVG_CHROMA_T ? state->tile->frame->cu_array : NULL,
+                    UVG_CHROMA_T);
             intra_search.pred_cu.type = CU_INTRA;
           } else  if (intra_search.pred_cu.intra.mip_flag) {
             intra_mode = 0;
@@ -1346,9 +1347,20 @@ static double search_cu(
     fwrite(&state->search_cabac.ctx, 1,  sizeof(state->search_cabac.ctx), state->encoder_control->cabac_debug_file);
   }
 
+  bool can_split[6];
+  uvg_get_possible_splits(state, cu_loc, split_tree, tree_type, can_split);
+  can_split_cu &= can_split[1] || can_split[2] || can_split[3] || can_split[4] || can_split[5];
+
   // Recursively split all the way to max search depth.
   if (can_split_cu) {
-    const int split_type = depth == 2 ? TT_VER_SPLIT : QT_SPLIT;
+    lcu_t * split_lcu = MALLOC(lcu_t, 5);
+    enum split_type best_split = 0;
+    double best_split_cost = MAX_DOUBLE;
+    cabac_data_t post_seach_cabac;
+    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) continue;
       split_tree_t new_split = {
         split_tree.split_tree | split_type << (split_tree.current_depth * 3),
         split_tree.current_depth + 1,
@@ -1358,13 +1370,10 @@ static double search_cu(
     
       double split_cost = 0.0;
       int cbf = cbf_is_set_any(cur_cu->cbf);
-    cabac_data_t post_seach_cabac;
-    memcpy(&post_seach_cabac, &state->search_cabac, sizeof(post_seach_cabac));
       memcpy(&state->search_cabac, &pre_search_cabac, sizeof(post_seach_cabac));
 
 
       double split_bits = 0;
-    lcu_t split_lcu;
 
       if (cur_cu->log2_height + cur_cu->log2_width > 4) {
 
@@ -1387,13 +1396,15 @@ static double search_cu(
             above_cu = uvg_cu_array_at_const(state->tile->frame->chroma_cu_array, x >> 1, (y >> 1) - 1);
           }
         }
+        split_tree_t count_tree = split_tree;
+        count_tree.split_tree = split_tree.split_tree | split_type << (split_tree.current_depth * 3);
         uvg_write_split_flag(
           state,
           &state->search_cabac,
           left_cu,
           above_cu, 
           tree_type != UVG_CHROMA_T ? cu_loc : &separate_tree_chroma_loc,
-        split_tree,
+          count_tree,
           tree_type,
           &split_bits);
       }
@@ -1410,16 +1421,16 @@ static double search_cu(
         cu_loc_t new_cu_loc[4];
         uint8_t separate_chroma = 0;
         const int splits = uvg_get_split_locs(cu_loc, split_type, new_cu_loc, &separate_chroma);
-      initialize_partial_work_tree(lcu, &split_lcu, cu_loc, tree_type);
+        initialize_partial_work_tree(lcu, &split_lcu[split_type - 1], cu_loc, tree_type);
         for (int split = 0; split < splits; ++split) {
           new_split.part_index = split;
           split_cost += search_cu(state, 
             &new_cu_loc[split], separate_chroma ? chroma_loc : &new_cu_loc[split],
-          &split_lcu, 
+            &split_lcu[split_type -1], 
             tree_type, new_split,
             !separate_chroma || split == splits - 1);
           // If there is no separate chroma the block will always have chroma, otherwise it is the last block of the split that has the chroma
-        if (split_cost > cost) {
+          if (split_cost > cost || split_cost > best_split_cost) {
             break;
           }
         }
@@ -1427,6 +1438,12 @@ static double search_cu(
       } else {
         split_cost = INT_MAX;
       }
+      if (split_cost < best_split_cost) {
+        best_split_cost = split_cost;
+        best_split = split_type;
+        memcpy(&best_split_cabac, &state->search_cabac, sizeof(cabac_data_t));
+      }
+    }
 
     // If no search is not performed for this depth, try just the best mode
     // of the top left CU from the next depth. This should ensure that 64x64
@@ -1440,7 +1457,7 @@ static double search_cu(
       && tree_type == UVG_BOTH_T)
     {
 
-      cu_info_t *cu_d1 = LCU_GET_CU_AT_PX(&split_lcu, x_local, y_local);
+      cu_info_t *cu_d1 = LCU_GET_CU_AT_PX(&split_lcu[best_split - 1], x_local, y_local);
 
       // If the best CU in depth+1 is intra and the biggest it can be, try it.
       if (cu_d1->type == CU_INTRA && (cu_d1->log2_height + 1 == cur_cu->log2_height || cu_d1->log2_width + 1 == cur_cu->log2_width)) {
@@ -1488,10 +1505,14 @@ static double search_cu(
       }
     }
 
-    if (split_cost < cost) {
+    if (best_split_cost < cost) {
       // Copy split modes to this depth.
-      cost = split_cost;
+      cost = best_split_cost;
-      work_tree_copy_up(&split_lcu, lcu, state->encoder_control->cfg.jccr, tree_type, cu_loc, is_separate_tree && !has_chroma ? NULL : chroma_loc);
+      memcpy(&state->search_cabac, &best_split_cabac, sizeof(best_split_cabac));
+      work_tree_copy_up(&split_lcu[best_split -1], lcu, state->encoder_control->cfg.jccr, tree_type, cu_loc, is_separate_tree && !has_chroma ? NULL : chroma_loc);
+      downsample_cclm_rec(
+        state, x, y, cu_width / 2, cu_height / 2, lcu->rec.y, lcu->left_ref.y[64]
+      );
 #if UVG_DEBUG
       //debug_split = 1;
 #endif
@@ -1522,6 +1543,7 @@ static double search_cu(
         state, x, y, cu_width / 2, cu_height / 2, lcu->rec.y, lcu->left_ref.y[64]
       );      
     }
+    FREE_POINTER(split_lcu);
   } else if (cur_cu->log2_height + cur_cu->log2_width > 4) {
     // Need to copy modes down since the lower level of the work tree is used
     // when searching SMP and AMP blocks.
@@ -1745,11 +1767,11 @@ void uvg_search_lcu(encoder_state_t * const state, const int x, const int y, con
   double cost = search_cu(
     state, 
     &start,
-    NULL,
+    &start,
     &work_tree,
     tree_type,
     split_tree,
-    false);
+    true);
 
   // Save squared cost for rate control.
   if(state->encoder_control->cfg.rc_algorithm == UVG_LAMBDA) {
@@ -1766,10 +1788,10 @@ void uvg_search_lcu(encoder_state_t * const state, const int x, const int y, con
   if(state->frame->slicetype == UVG_SLICE_I && state->encoder_control->cfg.dual_tree) {
     cost = search_cu(
       state, &start,
-      NULL,
+      &start,
       &work_tree, UVG_CHROMA_T,
       split_tree,
-      false);
+      true);
 
     if (state->encoder_control->cfg.rc_algorithm == UVG_LAMBDA) {
       uvg_get_lcu_stats(state, x / LCU_WIDTH, y / LCU_WIDTH)->weight += cost * cost;