[intra] Extend the reference inside the reference building

Author: Jovasa

src/intra.c

@@ -198,8 +198,8 @@ static void intra_filter_reference(
     refs->filtered_initialized = true;
   }
 
-  const int_fast8_t ref_width = 2 * (1 << log2_width) + 1;
-  const int_fast8_t ref_height = 2 * (1 << log2_height) + 1;
+  const int_fast16_t ref_width = 2 * (1 << (log2_width)) + 1;
+  const int_fast16_t ref_height = 2 * (1 << (log2_height)) + 1;
   uvg_intra_ref *ref = &refs->ref;
   uvg_intra_ref *filtered_ref = &refs->filtered_ref;
 
@@ -208,15 +208,15 @@ static void intra_filter_reference(
   filtered_ref->top[0] = filtered_ref->left[0];
 
   // Top to bottom
-  for (int_fast8_t y = 1; y < ref_height - 1; ++y) {
+  for (int_fast16_t y = 1; y < ref_height - 1; ++y) {
     uvg_pixel *p = &ref->left[y];
     filtered_ref->left[y] = (p[-1] + 2 * p[0] + p[1] + 2) >> 2;
   }
   // Bottom left (not filtered) 
   filtered_ref->left[ref_height - 1] = ref->left[ref_height - 1];
 
   // Left to right
-  for (int_fast8_t x = 1; x < ref_width - 1; ++x) {
+  for (int_fast16_t x = 1; x < ref_width - 1; ++x) {
     uvg_pixel *p = &ref->top[x];
     filtered_ref->top[x] = (p[-1] + 2 * p[0] + p[1] + 2) >> 2;
   }
@@ -712,7 +712,16 @@ static void intra_predict_regular(
   }
 
   if (used_ref == &refs->filtered_ref && !refs->filtered_initialized) {
-    intra_filter_reference(log2_width, log2_height, refs);
+    int temp_log2_width = log2_width;
+    int temp_log2_height = log2_height;
+    if (color == COLOR_Y && isp_mode == ISP_MODE_NO_ISP) {
+      temp_log2_width = cur_cu->log2_width;
+      temp_log2_height = cur_cu->log2_height;
+    } else if (color != COLOR_Y) {
+      temp_log2_width = cur_cu->log2_chroma_width;
+      temp_log2_height = cur_cu->log2_chroma_height;
+    }
+    intra_filter_reference(temp_log2_width, temp_log2_height, refs);
   }
 
   if (mode == 0) {
@@ -797,7 +806,7 @@ void uvg_intra_build_reference_any(
   // Init pointers to LCUs reconstruction buffers, such that index 0 refers to block coordinate 0.
   const uvg_pixel *left_ref;
   bool extra_ref = false;
-  // On the left LCU edge, if left neighboring LCU is available, 
+  // On the left LCU edge, if left neighboring LCU is available,
   // left_ref needs to point to correct extra reference line if MRL is used.
   if (luma_px->x > 0 && lcu_px.x == 0 && multi_ref_index != 0) {
     left_ref = &extra_ref_lines[multi_ref_index * 128];
@@ -836,7 +845,7 @@ void uvg_intra_build_reference_any(
 
   const int log2_ratio    = log2_width - log2_height;
   int       s             = MAX(0, -log2_ratio);
-  int       mrl_extension = (multi_ref_index << s) + 3;
+  int       mrl_extension = (multi_ref_index << s) + (height << s) + 2;
   // Generate left reference.
   if (luma_px->x > 0) {
     // Get the number of reference pixels based on the PU coordinate within the LCU.
@@ -870,15 +879,17 @@ void uvg_intra_build_reference_any(
 
     // If first isp split, take samples as if it were normal square block
     int tmp_h = is_first_isp_block ? cu_height * 2 : (isp_mode ? cu_height + height : height * 2);
-    for (int i = px_available_left; i < tmp_h + mrl_extension; ++i) {
+    int total_height = MIN(tmp_h + mrl_extension, INTRA_REF_LENGTH);
+    for (int i = px_available_left; i < total_height; ++i) {
       out_left_ref[i + 1 + multi_ref_index] = nearest_pixel;
     }
   } else {
     // If we are on the left edge, extend the first pixel of the top row.
     uvg_pixel nearest_pixel = luma_px->y > 0 ? top_border[0] : dc_val;
     // If first isp split, take samples as if it were normal square block
     int tmp_h = is_first_isp_block ? cu_height * 2 : (isp_mode ? cu_height + height : height * 2);
-    for (int i = 0; i < tmp_h +  mrl_extension; i++) {
+    int total_height = MIN(tmp_h + mrl_extension, INTRA_REF_LENGTH);
+    for (int i = 0; i < total_height; i++) {
       // Reserve space for top left reference
       out_left_ref[i + 1 + multi_ref_index] = nearest_pixel;
     }
@@ -901,11 +912,11 @@ void uvg_intra_build_reference_any(
         // LCU left border case
         uvg_pixel *top_left_corner = &extra_ref_lines[multi_ref_index * 128];
         switch (multi_ref_index) {
-          case 0:          
+          case 0:
             out_left_ref[0] = left_border[(-1) * left_stride];
             out_top_ref[0] = top_left_corner[MAX_REF_LINE_IDX - 1];
             break;
-          case 1:          
+          case 1:
             for (int i = 0; i <= 1; ++i) {
               out_left_ref[i] = left_border[(i - 1 - 1) * left_stride];
               out_top_ref[i] = top_left_corner[(128 * -i) + MAX_REF_LINE_IDX - 1 - 1];
@@ -1002,7 +1013,7 @@ void uvg_intra_build_reference_any(
   // Generate top reference.
   int px_available_top;
   s             = MAX(0, log2_ratio);
-  mrl_extension = (multi_ref_index << s) + 3;
+  mrl_extension = (multi_ref_index << s) + (width << s) + 2;
   if (luma_px->y > 0) {
     // Get the number of reference pixels based on the PU coordinate within the LCU.
     if (isp_mode && !is_first_isp_block && !is_chroma) {
@@ -1018,7 +1029,7 @@ void uvg_intra_build_reference_any(
       const int num_cus = uvg_count_available_edge_cus(cu_loc, lcu, false);
       px_available_top = !is_chroma ? num_cus * 4 : num_cus * 2;
     }
-    
+
     // Limit the number of available pixels based on block size and dimensions
     // of the picture.
     px_available_top = MIN(px_available_top, cu_width + pu_loc->width);
@@ -1033,7 +1044,8 @@ void uvg_intra_build_reference_any(
 
     // If first isp split, take samples as if it were normal square block
     int tmp_w = is_first_isp_block ? cu_width * 2 : (isp_mode ? cu_width + width : width * 2);
-    for (int i = px_available_top; i < tmp_w +  mrl_extension; ++i) {
+    int total_width = MIN(tmp_w + mrl_extension, INTRA_REF_LENGTH);
+    for (int i = px_available_top; i < total_width; ++i) {
       out_top_ref[i + 1 + multi_ref_index] = nearest_pixel;
     }
   } else {
@@ -1042,7 +1054,8 @@ void uvg_intra_build_reference_any(
 
     // If first isp split, take samples as if it were normal square block
     int tmp_w = is_first_isp_block ? cu_width * 2 : (isp_mode ? cu_width + width : width * 2);
-    for (int i = 0; i < tmp_w + mrl_extension; i++) {
+    int total_width = MIN(tmp_w + mrl_extension, INTRA_REF_LENGTH);
+    for (int i = 0; i < total_width; i++) {
       out_top_ref[i + 1] = nearest_pixel;
     }
   }
@@ -1266,9 +1279,10 @@ void uvg_intra_build_reference_inner(
   // If first isp split, take samples as if it were normal square block
   const int log2_ratio    = log2_width - log2_height;
   int       s             = MAX(0, -log2_ratio);
-  int       mrl_extension = (multi_ref_index << s) + 3;
+  int       mrl_extension = ((multi_ref_index + 0) << s) + (height << s) + 2;
   int tmp_h = is_first_isp_block ? cu_height * 2 : (isp_mode ? cu_height + height : height * 2);
-  for (; i < tmp_h + mrl_extension; i += 4) {
+  int total_height = MIN(tmp_h + mrl_extension, INTRA_REF_LENGTH - 2);
+  for (; i < total_height; i += 4) {
     out_left_ref[i + 1] = nearest_pixel;
     out_left_ref[i + 2] = nearest_pixel;
     out_left_ref[i + 3] = nearest_pixel;
@@ -1314,9 +1328,10 @@ void uvg_intra_build_reference_inner(
 
   // If first isp split, take samples as if it were normal square block
   s             = MAX(0, -log2_ratio);
-  mrl_extension = (multi_ref_index << s) + 3;
+  mrl_extension = ((multi_ref_index + 0) << s) + (width << s) + 2;
   int tmp_w = is_first_isp_block ? cu_width * 2 : (isp_mode ? cu_width + width : width * 2);
-  for (; i < tmp_w + mrl_extension; i += 4) {
+  int total_width = MIN(tmp_w+ mrl_extension, INTRA_REF_LENGTH - 2);
+  for (; i < total_width; i += 4) {
     out_top_ref[i + 1 + multi_ref_index] = nearest_pixel;
     out_top_ref[i + 2 + multi_ref_index] = nearest_pixel;
     out_top_ref[i + 3 + multi_ref_index] = nearest_pixel;