diff --git a/src/encoder.c b/src/encoder.c
index 2ad4ab95..2376b1c2 100644
--- a/src/encoder.c
+++ b/src/encoder.c
@@ -1112,7 +1112,15 @@ void encode_coding_tree(encoder_control *encoder, uint16_t x_ctb,
             ti.cb_top[2] = (ti.cb[0] & 0x4 || ti.cb[1] & 0x4 || ti.cb[2] & 0x4 || ti.cb[3] & 0x4)?1:0;
 
             residual = ti.cb_top[0] | ti.cb_top[1] | ti.cb_top[2];
-            picture_set_block_residual(encoder->in.cur_pic,x_ctb,y_ctb,depth,ti.cb_top[0]);
+            if(depth == 0)  {
+              picture_set_block_residual(encoder->in.cur_pic,x_ctb    ,y_ctb    ,depth+1,ti.cb[0] & 0x1);
+              picture_set_block_residual(encoder->in.cur_pic,x_ctb + 4,y_ctb    ,depth+1,ti.cb[1] & 0x1);
+              picture_set_block_residual(encoder->in.cur_pic,x_ctb    ,y_ctb + 4,depth+1,ti.cb[2] & 0x1);
+              picture_set_block_residual(encoder->in.cur_pic,x_ctb + 4,y_ctb + 4,depth+1,ti.cb[3] & 0x1);
+            } else  {
+              picture_set_block_residual(encoder->in.cur_pic,x_ctb,y_ctb,depth,ti.cb_top[0]);
+            }
+            
 
             cabac.ctx = &g_cu_qt_root_cbf_model;
             CABAC_BIN(&cabac, residual, "rqt_root_cbf");
diff --git a/src/filter.c b/src/filter.c
index fe125c88..2d084812 100644
--- a/src/filter.c
+++ b/src/filter.c
@@ -190,21 +190,20 @@ void filter_deblock_edge_luma(encoder_control *encoder,
       if((block_idx & 1) == 0)
       {
         // CU in the side we are filtering, update every 8-pixels
-        cu_p = &encoder->in.cur_pic->cu_array[MAX_DEPTH][(x_cu - (dir == EDGE_VER) + (dir == EDGE_HOR ? block_idx/2 : 0)) +
-                                                         (y_cu - (dir == EDGE_HOR) + (dir == EDGE_VER ? block_idx/2 : 0))
+        cu_p = &encoder->in.cur_pic->cu_array[MAX_DEPTH][(x_cu - (dir == EDGE_VER) + (dir == EDGE_HOR ? block_idx>>1 : 0)) +
+                                                         (y_cu - (dir == EDGE_HOR) + (dir == EDGE_VER ? block_idx>>1 : 0))
                                                           * (encoder->in.width_in_lcu << MAX_DEPTH)];
         // Filter strength
-        strength = 0;
+        strength = 0;        
         // Intra blocks have strength 2
         if(cu_q->type == CU_INTRA || cu_p->type == CU_INTRA) {
           strength = 2;          
           // Non-zero residual and transform boundary
-        } else if((cu_q->residual || cu_p->residual) && (cu_q->depth==0 ? !((dir == EDGE_VER ? y_cu + (block_idx>>1) : x_cu + (block_idx>>1) )&0x3) : 1 ) ) {
+        } else if(cu_q->residual || cu_p->residual) {
           strength = 1;
           // Absolute motion vector diff between blocks >= 1 (Integer pixel)
         } else if((abs(cu_q->inter.mv[0] - cu_p->inter.mv[0]) >= 4) || (abs(cu_q->inter.mv[1] - cu_p->inter.mv[1]) >= 4)) {
-          strength = 1;
-         
+          strength = 1;         
         }
         tc_index        = CLIP(0, 51 + 2, (int32_t)(qp + 2*(strength - 1) + (tc_offset_div2 << 1)));
         tc              = g_tc_table_8x8[tc_index] * bitdepth_scale;