[ibc] Fill the IBC hashmap at the start of LCU search and use reverse map for "pos to hash"

src/encoder_state-ctors_dtors.c

@@ -133,6 +133,10 @@ static int encoder_state_config_tile_init(encoder_state_t * const state,
     state->tile->frame->ibc_buffer_v = malloc(sizeof(uvg_pixel*) * state->tile->frame->height_in_lcu);
     state->tile->frame->ibc_hashmap_row = malloc(sizeof(uvg_hashmap_t) * state->tile->frame->height_in_lcu);
 
+    state->tile->frame->ibc_hashmap_pos_to_hash_stride = ((state->tile->frame->width+UVG_HASHMAP_BLOCKSIZE-1)/ UVG_HASHMAP_BLOCKSIZE);
+    state->tile->frame->ibc_hashmap_pos_to_hash = malloc(sizeof(uint32_t) *
+      ((state->tile->frame->height+UVG_HASHMAP_BLOCKSIZE-1)/ UVG_HASHMAP_BLOCKSIZE) * state->tile->frame->ibc_hashmap_pos_to_hash_stride);
+
     for (uint32_t i = 0; i < state->tile->frame->height_in_lcu; i++) {
       state->tile->frame->ibc_hashmap_row[i] = uvg_hashmap_create((LCU_WIDTH * IBC_BUFFER_WIDTH)>>2);
       state->tile->frame->ibc_buffer_y[i] = (uvg_pixel*)malloc(IBC_BUFFER_SIZE * 3); // ToDo: we don't need this much, but it would also support 4:4:4
@@ -220,6 +224,8 @@ static void encoder_state_config_tile_finalize(encoder_state_t * const state) {
   FREE_POINTER(state->tile->frame->hmvp_size_ibc);
 
   if (state->encoder_control->cfg.ibc) {
+    FREE_POINTER(state->tile->frame->ibc_hashmap_pos_to_hash);
+
     for (uint32_t i = 0; i < state->tile->frame->height_in_lcu; i++) {
       FREE_POINTER(state->tile->frame->ibc_buffer_y[i]);
       uvg_hashmap_free(state->tile->frame->ibc_hashmap_row[i]);

src/encoderstate.c

@@ -288,23 +288,6 @@ static void encoder_state_recdata_to_bufs(encoder_state_t * const state,
     const uint32_t ibc_block_width = MIN(LCU_WIDTH, (state->tile->frame->width-lcu->position_px.x));
     const uint32_t ibc_block_height = MIN(LCU_WIDTH, (state->tile->frame->height-lcu->position_px.y));
 
-    int items = 0;
-    // Hash the current LCU to the IBC hashmap
-    for (int32_t xx = (lcu->position_px.x>8)?-6:0; xx < (int32_t)(ibc_block_width)-7; xx+=2) {
-      for (int32_t yy = 0; yy < (int32_t)(ibc_block_height)-7; yy++) {
-        int cur_x = lcu->position_px.x + xx;
-        int cur_y = lcu->position_px.y + yy;
-        uint32_t crc = uvg_crc32c_8x8(&frame->rec->y[cur_y * frame->rec->stride + cur_x],frame->rec->stride);
-        if (state->encoder_control->chroma_format != UVG_CSP_400) {
-          crc ^= uvg_crc32c_4x4(&frame->rec->u[(cur_y>>1) * (frame->rec->stride>>1) + (cur_x>>1)],frame->rec->stride>>1);
-          crc ^= uvg_crc32c_4x4(&frame->rec->v[(cur_y>>1) * (frame->rec->stride>>1) + (cur_x>>1)],frame->rec->stride>>1);
-        }
-        uvg_hashmap_insert(frame->ibc_hashmap_row[ibc_buffer_row], crc, ((cur_x&0xffff)<<16) | (cur_y&0xffff));
-        items++;
-      }
-    }
-    //fprintf(stderr, "Inserted %d items to %dx%d at %dx%d\r\n", items, ibc_block_width, ibc_block_height, lcu->position_px.x, lcu->position_px.y);
-
     uvg_pixels_blit(&frame->rec->y[lcu->position_px.y * frame->rec->stride + lcu->position_px.x],
                     &frame->ibc_buffer_y[ibc_buffer_row][ibc_buffer_pos_x],
                     ibc_block_width, ibc_block_height,
@@ -771,6 +754,49 @@ static void encoder_state_worker_encode_lcu_search(void * opaque)
   if(state->frame->slicetype != UVG_SLICE_I) memcpy(original_lut, &state->tile->frame->hmvp_lut[ctu_row_mul_five], sizeof(cu_info_t) * MAX_NUM_HMVP_CANDS);
   if(state->encoder_control->cfg.ibc) memcpy(original_lut_ibc, &state->tile->frame->hmvp_lut_ibc[ctu_row_mul_five], sizeof(cu_info_t) * MAX_NUM_HMVP_CANDS);
 
+
+  if (state->encoder_control->cfg.ibc) {
+    videoframe_t * const frame      = state->tile->frame;
+    const uint32_t ibc_block_width  = MIN(LCU_WIDTH, (state->tile->frame->width-lcu->position_px.x));
+    const uint32_t ibc_block_height = MIN(LCU_WIDTH, (state->tile->frame->height-lcu->position_px.y));
+    int items = 0;
+    // Hash the current LCU to the IBC hashmap
+    for (int32_t xx = (lcu->position_px.x>8)?-7:0; xx < (int32_t)(ibc_block_width)-7; xx++) {
+      for (int32_t yy = 0; yy < (int32_t)(ibc_block_height)-7; yy++) {
+        int cur_x = lcu->position_px.x + xx;
+        int cur_y = lcu->position_px.y + yy;
+        
+        // Skip blocks that seem to be the same value for the whole block
+        uint64_t first_line =
+          *(uint64_t *)&frame->source->y[cur_y * frame->source->stride + cur_x];
+        bool same_data = true;
+        for (int y_temp = 1; y_temp < 8; y_temp++) {
+          if (*(uint64_t *)&frame->source->y[(cur_y+y_temp) * frame->source->stride + cur_x] != first_line) {
+            same_data = false;
+            break;
+          }
+        }
+        
+        if (!same_data || (xx % UVG_HASHMAP_BLOCKSIZE == 0 && yy % UVG_HASHMAP_BLOCKSIZE == 0)) {
+          uint32_t crc = uvg_crc32c_8x8(&frame->source->y[cur_y * frame->source->stride + cur_x],frame->source->stride);       
+          if (xx % UVG_HASHMAP_BLOCKSIZE == 0 && yy % UVG_HASHMAP_BLOCKSIZE == 0) {
+            state->tile->frame->ibc_hashmap_pos_to_hash[(cur_y / UVG_HASHMAP_BLOCKSIZE)*state->tile->frame->ibc_hashmap_pos_to_hash_stride + cur_x / UVG_HASHMAP_BLOCKSIZE] = crc;
+          }
+          /*
+          if (state->encoder_control->chroma_format != UVG_CSP_400) {
+            crc ^= uvg_crc32c_4x4(&frame->rec->u[(cur_y>>1) * (frame->rec->stride>>1) + (cur_x>>1)],frame->rec->stride>>1);
+            crc ^= uvg_crc32c_4x4(&frame->rec->v[(cur_y>>1) * (frame->rec->stride>>1) + (cur_x>>1)],frame->rec->stride>>1);
+          }
+          */
+          uvg_hashmap_insert(frame->ibc_hashmap_row[ctu_row], crc, ((cur_x&0xffff)<<16) | (cur_y&0xffff));
+          items++;
+        }
+      }
+    }
+  }
+  //fprintf(stderr, "Inserted %d items to %dx%d at %dx%d\r\n", items, ibc_block_width, ibc_block_height, lcu->position_px.x, lcu->position_px.y);
+
+
   //This part doesn't write to bitstream, it's only search, deblock and sao
   uvg_search_lcu(state, lcu->position_px.x, lcu->position_px.y, state->tile->hor_buf_search, state->tile->ver_buf_search, lcu->coeff);
 

src/hashmap.c

@@ -75,9 +75,9 @@ uvg_hashmap_t* uvg_hashmap_create(uint32_t bucket_size)
  */
 static uint32_t uvg_hashmap_hash(uint32_t key, uint32_t bucket_size)
 {
-  key ^= (key >> 20) ^ (key >> 12);
+  //key ^= (key >> 20) ^ (key >> 12);
-  return (key ^ (key >> 7) ^ (key >> 4) ^ 2654435769U) % bucket_size;
+  //return (key ^ (key >> 7) ^ (key >> 4) ^ 2654435769U) % bucket_size;
-  //return key % bucket_size;
+  return key % bucket_size;
 }
 
 /**

src/search.c

@@ -1012,6 +1012,7 @@ static double search_cu(
     // Simple IBC search
     if (can_use_intra //&& state->frame->slicetype == UVG_SLICE_I
          && state->encoder_control->cfg.ibc 
+         && cost > 1000
          && cu_width > 4
          && (x >= cu_width || y >= cu_width)
          && !cur_cu->skipped) {
@@ -1029,6 +1030,7 @@ static double search_cu(
         cost = mode_cost;
         inter_bitcost = mode_bitcost;
         cur_cu->type = CU_IBC;
+        cur_cu->inter.mv_dir = 1;
         cur_cu->joint_cb_cr = 0;
       } else {
         *cur_cu = backup_cu;

src/search_ibc.c

@@ -1083,6 +1083,7 @@ static int uvg_search_hash_cu_ibc(encoder_state_t* const state,
     lcu);
 
   *inter_cost               = MAX_DOUBLE;
+  *inter_bitcost            = MAX_DOUBLE;
 
   bool valid_mv = false;
   
@@ -1112,11 +1113,12 @@ static int uvg_search_hash_cu_ibc(encoder_state_t* const state,
   uint32_t      ibc_buffer_row = yy / LCU_WIDTH;
 
   //UVG_GET_TIME(&hashmap_start_temp);
-  uint32_t crc = uvg_crc32c_8x8(&state->tile->frame->source->y[yy * state->tile->frame->source->stride + xx],state->tile->frame->source->stride);
+  uint32_t crc = state->tile->frame->ibc_hashmap_pos_to_hash[(yy / UVG_HASHMAP_BLOCKSIZE)*state->tile->frame->ibc_hashmap_pos_to_hash_stride + xx / UVG_HASHMAP_BLOCKSIZE];
-  if (state->encoder_control->chroma_format != UVG_CSP_400) {
+  //uvg_crc32c_8x8(&state->tile->frame->source->y[yy * state->tile->frame->source->stride + xx],state->tile->frame->source->stride);
+  /* if (state->encoder_control->chroma_format != UVG_CSP_400) {
     crc ^= uvg_crc32c_4x4(&state->tile->frame->source->u[(yy >> 1) * (state->tile->frame->source->stride>>1) + (xx >> 1)],state->tile->frame->source->stride>>1);
     crc ^= uvg_crc32c_4x4(&state->tile->frame->source->v[(yy >> 1) * (state->tile->frame->source->stride>>1) + (xx >> 1)],state->tile->frame->source->stride>>1);
-  }
+  }*/
   /* UVG_GET_TIME(&hashmap_end_temp);
   crc_time += UVG_CLOCK_T_AS_DOUBLE(hashmap_end_temp) -
                 UVG_CLOCK_T_AS_DOUBLE(hashmap_start_temp);*/
@@ -1133,7 +1135,7 @@ static int uvg_search_hash_cu_ibc(encoder_state_t* const state,
 
   while (result != NULL) {
     if (hashes_found == 0 && result->size > 1000) {
-      fprintf(stderr, "Found a block with %d elements\n", result->size);
+      //fprintf(stderr, "Found a block with %d elements\n", result->size);
       //break;      
     }
     if (result->key == crc && result->value != own_location) {
@@ -1149,11 +1151,13 @@ static int uvg_search_hash_cu_ibc(encoder_state_t* const state,
           bool full_block = true; // Is the full block covered by the IBC?
           for (int xxx = xx+UVG_HASHMAP_BLOCKSIZE; xxx < xx + width; xxx+=UVG_HASHMAP_BLOCKSIZE) {
             for (int yyy = yy; yyy < yy + height; yyy += UVG_HASHMAP_BLOCKSIZE) {
-              uint32_t crc_other_blocks = uvg_crc32c_8x8(&state->tile->frame->source->y[yyy * state->tile->frame->source->stride + xxx],state->tile->frame->source->stride);
+              uint32_t crc_other_blocks = state->tile->frame->ibc_hashmap_pos_to_hash[(yyy / UVG_HASHMAP_BLOCKSIZE)*state->tile->frame->ibc_hashmap_pos_to_hash_stride + xxx / UVG_HASHMAP_BLOCKSIZE];
+              //uvg_crc32c_8x8(&state->tile->frame->source->y[yyy * state->tile->frame->source->stride + xxx],state->tile->frame->source->stride);
+              /*
               if (state->encoder_control->chroma_format != UVG_CSP_400) {
                 crc_other_blocks ^= uvg_crc32c_4x4(&state->tile->frame->source->u[(yyy >> 1) * (state->tile->frame->source->stride>>1) + (xxx >> 1)],state->tile->frame->source->stride>>1);
                 crc_other_blocks ^= uvg_crc32c_4x4(&state->tile->frame->source->v[(yyy >> 1) * (state->tile->frame->source->stride>>1) + (xxx >> 1)],state->tile->frame->source->stride>>1);
-              }
+              }*/
               uvg_hashmap_node_t *result2 = uvg_hashmap_search(state->tile->frame->ibc_hashmap_row[ibc_buffer_row],crc_other_blocks);
               evaluations++;
               bool found_match = false;
@@ -1180,19 +1184,22 @@ static int uvg_search_hash_cu_ibc(encoder_state_t* const state,
               break;
             } 
           }
+          double     cost = *inter_cost, bits = *inter_bitcost;
+          vector2d_t mv = { best_mv_x, best_mv_y};
           
-          if (full_block) {
+          if (full_block && check_mv_cost(&info, mv_x, mv_y, &cost, &bits, &mv)) {
 
-            double cost    = get_ibc_mvd_coding_cost(state, &state->cabac, mv_x,mv_y) * state->lambda_sqrt;
+            //double cost    = get_ibc_mvd_coding_cost(state, &state->cabac, mv_x,mv_y) * state->lambda_sqrt;
+            //cost += 
             bool better_mv = cost < *inter_cost;
             if (better_mv) {
               best_mv_x              = mv_x;
               best_mv_y              = mv_y;
               *inter_cost            = cost;
-              *inter_bitcost         = 0.0;
+              *inter_bitcost         = bits;
               fprintf(stderr, "Found best IBC!! %dx%d %dx%d: %d,%d\r\n", x,y, width,width, mv_x, mv_y);
               found_block = true;
-              break;
+              //break;
             }
           }
         }
@@ -1256,12 +1263,13 @@ void uvg_search_cu_ibc(encoder_state_t * const state,
   *inter_cost = MAX_DOUBLE;
   *inter_bitcost = MAX_INT;
    // Quick hashmap search
-  uvg_search_hash_cu_ibc(state,
+  /* uvg_search_hash_cu_ibc(
+    state,
                           x, y, depth,
                           lcu,
                           inter_cost,
                           inter_bitcost);
-  return;
+  return;*/
   // Store information of L0, L1, and bipredictions.
   // Best cost will be left at MAX_DOUBLE if no valid CU is found.
   // These will be initialized by the following function.

src/videoframe.h

@@ -83,6 +83,8 @@ typedef struct videoframe
   uvg_pixel **ibc_buffer_u; //!< \brief Intra Block Copy buffer for each LCU row 
   uvg_pixel **ibc_buffer_v; //!< \brief Intra Block Copy buffer for each LCU row
   uvg_hashmap_t **ibc_hashmap_row; //!< \brief Hashmap for IBC hash search for each LCU row
+  uint32_t *ibc_hashmap_pos_to_hash; //!< \brief Hashmap reverse search for position to hash
+  uint32_t ibc_hashmap_pos_to_hash_stride; //!< \brief Hashmap position to hash stride
   cu_info_t* hmvp_lut_ibc; //!< \brief Look-up table for HMVP in IBC, one for each LCU row
   uint8_t* hmvp_size_ibc; //!< \brief HMVP IBC LUT size