[cclm] fix cclm when deblocking is enabled

src/intra.c

@@ -497,17 +497,11 @@ void kvz_predict_cclm(
     for (; available_above_right < width / 2; available_above_right++) {
       int x_extension = x_scu + width * 2 + 4 * available_above_right;
       cu_info_t* pu = LCU_GET_CU_AT_PX(lcu, x_extension, y_scu - 4);
-      if (pu->type == CU_NOTSET || x_extension > LCU_WIDTH) break;
+      if (x_extension >= LCU_WIDTH || pu->type == CU_NOTSET) break;
     }
     if(y_scu == 0) {
       if(!state->encoder_control->cfg.wpp) available_above_right = MIN(width / 2, (state->tile->frame->width - x0 - width * 2) / 4);
-      for (int x = 0; x < width * (available_above_right ? 4 : 2); x += 2) {
+      memcpy(sampled_luma_ref.top, &state->tile->frame->cclm_luma_rec_top_line[x0 / 2 + (y0 / 64 - 1) * (stride / 2)], sizeof(kvz_pixel) * (width + available_above_right * 2));
-        bool left_padding = x0 || x;
-        sampled_luma_ref.top[x / 2] = (state->tile->frame->rec->y[x0 + x + (y0 - 1) * stride] * 2 +
-          state->tile->frame->rec->y[x0 + x + 1 + (y0 - 1) * stride] +
-          state->tile->frame->rec->y[x0 + x - left_padding + (y0 - 1) * stride] + 
-          2) >> 2;
-      }
     }
     else {
       for (int x = 0; x < width * (available_above_right ? 4 : 2); x += 2) {

src/search.c

@@ -241,7 +241,7 @@ static double cu_zero_coeff_cost(const encoder_state_t *state, lcu_t *work_tree,
 }
 
 
-static void downsample_cclm_rec(encoder_state_t *state, int x, int y, int width, int height, kvz_pixel *y_rec) {
+static void downsample_cclm_rec(encoder_state_t *state, int x, int y, int width, int height, kvz_pixel *y_rec, kvz_pixel extra_pixel) {
   if (!state->encoder_control->cfg.cclm) return;
   int x_scu = SUB_SCU(x);
   int y_scu = SUB_SCU(y);
@@ -265,6 +265,17 @@ static void downsample_cclm_rec(encoder_state_t *state, int x, int y, int width,
     }
     y_rec += LCU_WIDTH * 2;
   }
+  if((y + height * 2) % 64 == 0) {
+    int line = y / 64 * stride / 2;
+    y_rec -= LCU_WIDTH;
+    for (int i = 0; i < width; ++i) {
+      int s = 2;
+      s += y_rec[i * 2] * 2;
+      s += y_rec[i * 2 + 1];
+      s += !x_scu && !i && x ? extra_pixel : y_rec[i * 2 - ((i + x) > 0)] ;
+      state->tile->frame->cclm_luma_rec_top_line[i + x / 2 + line] = s >> 2;
+    }
+  }
 }
 
 
@@ -739,7 +750,7 @@ static double search_cu(encoder_state_t * const state, int x, int y, int depth,
                          NULL, NULL, lcu);
 
       downsample_cclm_rec(
-        state, x, y, cu_width / 2, cu_width / 2, lcu->rec.y
+        state, x, y, cu_width / 2, cu_width / 2, lcu->rec.y, lcu->left_ref.y[64]
       );
 
       // TODO: This heavily relies to square CUs
@@ -945,7 +956,7 @@ static double search_cu(encoder_state_t * const state, int x, int y, int depth,
       // search.
       work_tree_copy_down(x_local, y_local, depth, work_tree);
       downsample_cclm_rec(
-        state, x, y, cu_width / 2, cu_width / 2, lcu->rec.y
+        state, x, y, cu_width / 2, cu_width / 2, lcu->rec.y, lcu->left_ref.y[64]
       );
 
       if (state->frame->slicetype != KVZ_SLICE_I) {
@@ -960,7 +971,7 @@ static double search_cu(encoder_state_t * const state, int x, int y, int depth,
     // when searching SMP and AMP blocks.
     work_tree_copy_down(x_local, y_local, depth, work_tree);
     downsample_cclm_rec(
-      state, x, y, cu_width / 2, cu_width / 2, lcu->rec.y
+      state, x, y, cu_width / 2, cu_width / 2, lcu->rec.y, lcu->left_ref.y[64]
     );
 
     if (state->frame->slicetype != KVZ_SLICE_I) {

src/videoframe.c

@@ -62,6 +62,7 @@ videoframe_t * kvz_videoframe_alloc(int32_t width,
     if (cclm) {
       assert(chroma_format == KVZ_CSP_420);
       frame->cclm_luma_rec = MALLOC(kvz_pixel, (((width + 7) & ~7) + FRAME_PADDING_LUMA) * (((height + 7) & ~7) + FRAME_PADDING_LUMA) / 4);
+      frame->cclm_luma_rec_top_line = MALLOC(kvz_pixel, (((width + 7) & ~7) + FRAME_PADDING_LUMA) / 2 * CEILDIV(height, 64));
     }
   }
   
@@ -83,6 +84,9 @@ int kvz_videoframe_free(videoframe_t * const frame)
   if(frame->cclm_luma_rec) {
     FREE_POINTER(frame->cclm_luma_rec);
   }
+  if(frame->cclm_luma_rec_top_line) {
+    FREE_POINTER(frame->cclm_luma_rec_top_line);
+  }
 
   kvz_image_free(frame->source);
   frame->source = NULL;

src/videoframe.h

@@ -54,6 +54,7 @@ typedef struct videoframe
   kvz_picture *rec_lmcs;       //!< \brief LMCS mapped reconstructed image, if available, otherwise points to source.
 
   kvz_pixel *cclm_luma_rec;    //!< \brief buffer for the downsampled luma reconstruction for cclm
+  kvz_pixel *cclm_luma_rec_top_line;    //!< \brief buffer for the downsampled luma reconstruction for cclm
 
   uint8_t* lmcs_avg_processed; //!< \brief For each LCU, indicates if already calculated average of border pixels is available
   int32_t* lmcs_avg;           //!< \brief Average of LCU border pixels