diff --git a/src/cu.c b/src/cu.c
index 5ed7a140..7f9f8e94 100644
--- a/src/cu.c
+++ b/src/cu.c
@@ -125,6 +125,8 @@ cu_array_t * kvz_cu_array_alloc(const int width_in_scu, const int height_in_scu)
   cu_array_t *cua;
   cua = MALLOC(cu_array_t, 1);
   cua->data = (cu_info_t*)malloc(sizeof(cu_info_t) * cu_array_size);
+  cua->width = width_in_scu << 3;
+  cua->height = height_in_scu << 3;
   cua->refcount = 1;
   FILL_ARRAY(cua->data, 0, cu_array_size);
   return cua;
@@ -145,3 +147,37 @@ int kvz_cu_array_free(cu_array_t * const cua)
   return 1;
 }
 
+
+/**
+ * \brief Copy part of a cu array to another cu array.
+ *
+ * All values are in luma pixels.
+ *
+ * \param dst     destination array
+ * \param dst_x   x-coordinate of the left edge of the copied area in dst
+ * \param dst_y   y-coordinate of the top edge of the copied area in dst
+ * \param src     source array
+ * \param src_x   x-coordinate of the left edge of the copied area in src
+ * \param src_y   y-coordinate of the top edge of the copied area in src
+ * \param width   width of the area to copy
+ * \param height  height of the area to copy
+ */
+void kvz_cu_array_copy(cu_array_t* dst,       int dst_x, int dst_y,
+                       const cu_array_t* src, int src_x, int src_y,
+                       int width, int height)
+{
+  // Convert values from pixel coordinates to array indices.
+  int src_stride = src->width >> 3;
+  int dst_stride = dst->width >> 3;
+  const cu_info_t* src_ptr = &src->data[(src_x >> 3) + (src_y >> 3) * src_stride];
+  cu_info_t* dst_ptr       = &dst->data[(dst_x >> 3) + (dst_y >> 3) * dst_stride];
+
+  // Number of bytes to copy per row.
+  const size_t row_size = sizeof(cu_info_t) * (width >> 3);
+
+  for (int i = 0; i < (height >> 3); ++i) {
+    memcpy(dst_ptr, src_ptr, row_size);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+  }
+}
diff --git a/src/cu.h b/src/cu.h
index 1517ef01..36f4a310 100644
--- a/src/cu.h
+++ b/src/cu.h
@@ -178,12 +178,18 @@ typedef struct
   (cu).inter.mv_cand, (cu).inter.mv_ref, (cu).inter.mv_dir, (cu).inter.mode)
 
 typedef struct {
-  cu_info_t *data;           //!< \brief cu_info data
-  int32_t refcount;        //!< \brief number of references in reflists to this cu_array
+  cu_info_t *data; //!< \brief cu array
+  int32_t width;    //!< \brief width of the array in pixels
+  int32_t height;   //!< \brief height of the array in pixels
+  int32_t refcount; //!< \brief number of references to this cu_array
 } cu_array_t;
 
 cu_array_t * kvz_cu_array_alloc(int width_in_scu, int height_in_scu);
 int kvz_cu_array_free(cu_array_t *cua);
+void kvz_cu_array_copy(cu_array_t* dst,       int dst_x, int dst_y,
+                       const cu_array_t* src, int src_x, int src_y,
+                       int width, int height);
+
 
 /**
  * \brief Return the 7 lowest-order bits of the pixel coordinate.
diff --git a/src/encoderstate.c b/src/encoderstate.c
index 6ca153b1..417109fb 100644
--- a/src/encoderstate.c
+++ b/src/encoderstate.c
@@ -207,18 +207,15 @@ static void encoder_state_worker_encode_lcu(void * opaque) {
     while (main_state->parent) main_state = main_state->parent;
     assert(main_state != state);
 
-    unsigned child_width_in_scu = state->tile->frame->width_in_lcu << MAX_DEPTH;
-    unsigned main_width_in_scu = main_state->tile->frame->width_in_lcu << MAX_DEPTH;
-    unsigned tile_x = state->tile->lcu_offset_x << MAX_DEPTH;
-    unsigned tile_y = state->tile->lcu_offset_y << MAX_DEPTH;
-    unsigned x = lcu->position.x << MAX_DEPTH;
-    unsigned y = lcu->position.y << MAX_DEPTH;
-
-    for (unsigned lcu_row = 0; lcu_row < 8; ++lcu_row) {
-      cu_info_t *main_row = &main_state->tile->frame->cu_array->data[x + tile_x + (y + tile_y + lcu_row) * main_width_in_scu];
-      cu_info_t *child_row = &state->tile->frame->cu_array->data[x + (y + lcu_row) * child_width_in_scu];
-      memcpy(main_row, child_row, sizeof(cu_info_t) * 8);
-    }
+    const unsigned tile_x_px = state->tile->lcu_offset_x << LOG2_LCU_WIDTH;
+    const unsigned tile_y_px = state->tile->lcu_offset_y << LOG2_LCU_WIDTH;
+    const unsigned x_px = lcu->position_px.x;
+    const unsigned y_px = lcu->position_px.y;
+    kvz_cu_array_copy(main_state->tile->frame->cu_array,
+                      x_px + tile_x_px, y_px + tile_y_px,
+                      state->tile->frame->cu_array,
+                      x_px, y_px,
+                      LCU_WIDTH, LCU_WIDTH);
 
     PERFORMANCE_MEASURE_END(KVZ_PERF_FRAME, state->encoder_control->threadqueue, "type=copy_cuinfo,frame=%d,tile=%d", state->global->frame, state->tile->id);
   }