Merge "Change to do LPF in SB64 order" into experimental

1 files changed, 380 insertions, 167 deletions

diff --git a/vp9/common/vp9_loopfilter.c b/vp9/common/vp9_loopfilter.c
index 9a5087326..e09dfced7 100644
--- a/vp9/common/vp9_loopfilter.c
+++ b/vp9/common/vp9_loopfilter.c
@@ -194,25 +194,213 @@ static int mb_lf_skip(const MB_MODE_INFO *const mbmi) {
 static int sb_mb_lf_skip(const MODE_INFO *const mip0,
                          const MODE_INFO *const mip1) {
   const MB_MODE_INFO *mbmi0 = &mip0->mbmi;
-  const MB_MODE_INFO *mbmi1 = &mip0->mbmi;
+  const MB_MODE_INFO *mbmi1 = &mip1->mbmi;
   return mb_lf_skip(mbmi0) && mb_lf_skip(mbmi1) &&
-         (mbmi0->ref_frame == mbmi1->ref_frame) &&
-         (mbmi0->mv[mbmi0->ref_frame].as_int ==
-          mbmi1->mv[mbmi1->ref_frame].as_int) &&
-         mbmi0->ref_frame != INTRA_FRAME;
+         mbmi0->ref_frame != INTRA_FRAME &&
+         mbmi1->ref_frame != INTRA_FRAME;
 }
 
+static void lpf_mb(VP9_COMMON *cm, const MODE_INFO *mi,
+                   int do_left_mb_v, int do_above_mb_h,
+                   int do_left_mbuv_v, int do_above_mbuv_h,
+                   uint8_t *y_ptr, uint8_t *u_ptr, uint8_t *v_ptr,
+                   int y_stride, int uv_stride, int dering) {
+  loop_filter_info_n *lfi_n = &cm->lf_info;
+  struct loop_filter_info lfi;
+  const FRAME_TYPE frame_type = cm->frame_type;
+  int mode = mi->mbmi.mode;
+  int mode_index = lfi_n->mode_lf_lut[mode];
+  int seg = mi->mbmi.segment_id;
+  int ref_frame = mi->mbmi.ref_frame;
+  int filter_level = lfi_n->lvl[seg][ref_frame][mode_index];
+
+  if (filter_level) {
+    const int skip_lf = mb_lf_skip(&mi->mbmi);
+    const int tx_size = mi->mbmi.txfm_size;
+    if (cm->filter_type == NORMAL_LOOPFILTER) {
+      const int hev_index = lfi_n->hev_thr_lut[frame_type][filter_level];
+      lfi.mblim = lfi_n->mblim[filter_level];
+      lfi.blim = lfi_n->blim[filter_level];
+      lfi.lim = lfi_n->lim[filter_level];
+      lfi.hev_thr = lfi_n->hev_thr[hev_index];
+
+      if (do_above_mb_h) {
+        if (tx_size >= TX_16X16)
+          vp9_lpf_mbh_w(y_ptr,
+                        do_above_mbuv_h ? u_ptr : NULL,
+                        do_above_mbuv_h ? v_ptr : NULL,
+                        y_stride, uv_stride, &lfi);
+        else
+          vp9_loop_filter_mbh(y_ptr, u_ptr, v_ptr, y_stride, uv_stride, &lfi);
+      }
+
+      if (!skip_lf) {
+        if (tx_size >= TX_8X8) {
+          if (tx_size == TX_8X8 && (mode == I8X8_PRED || mode == SPLITMV))
+            vp9_loop_filter_bh8x8(y_ptr, u_ptr, v_ptr,
+                                  y_stride, uv_stride, &lfi);
+          else
+            vp9_loop_filter_bh8x8(y_ptr, NULL, NULL,
+                                  y_stride, uv_stride, &lfi);
+        } else {
+          vp9_loop_filter_bh(y_ptr, u_ptr, v_ptr,
+                             y_stride, uv_stride, &lfi);
+        }
+      }
+
+      if (do_left_mb_v) {
+        if (tx_size >= TX_16X16)
+          vp9_lpf_mbv_w(y_ptr,
+                        do_left_mbuv_v ? u_ptr : NULL,
+                        do_left_mbuv_v ? v_ptr : NULL,
+                        y_stride, uv_stride, &lfi);
+        else
+          vp9_loop_filter_mbv(y_ptr, u_ptr, v_ptr, y_stride, uv_stride, &lfi);
+      }
+
+      if (!skip_lf) {
+        if (tx_size >= TX_8X8) {
+          if (tx_size == TX_8X8 && (mode == I8X8_PRED || mode == SPLITMV))
+            vp9_loop_filter_bv8x8(y_ptr, u_ptr, v_ptr,
+                                  y_stride, uv_stride, &lfi);
+          else
+            vp9_loop_filter_bv8x8(y_ptr, NULL, NULL,
+                                  y_stride, uv_stride, &lfi);
+        } else {
+          vp9_loop_filter_bv(y_ptr, u_ptr, v_ptr,
+                             y_stride, uv_stride, &lfi);
+        }
+      }
+      if (dering) {
+#if CONFIG_LOOP_DERING
+        vp9_post_proc_down_and_across(y_ptr, y_ptr,
+          y_stride, y_stride,
+          16, 16, dering);
+        if (u_ptr && v_ptr) {
+          vp9_post_proc_down_and_across(u_ptr, u_ptr,
+            uv_stride, uv_stride,
+            8, 8, dering);
+          vp9_post_proc_down_and_across(v_ptr, v_ptr,
+            uv_stride, uv_stride,
+            8, 8, dering);
+        }
+#endif
+      }
+    } else {
+      // TODO(yaowu): simple loop filter
+    }
+  }
+}
+
+static void lpf_sb32(VP9_COMMON *cm, const MODE_INFO *mode_info_context,
+                     int mb_row, int mb_col,
+                     uint8_t *y_ptr, uint8_t *u_ptr, uint8_t *v_ptr,
+                     int y_stride, int uv_stride,
+                     int y_only, int dering) {
+    BLOCK_SIZE_TYPE sb_type = mode_info_context->mbmi.sb_type;
+    TX_SIZE tx_size = mode_info_context->mbmi.txfm_size;
+    int do_left_v, do_above_h;
+    int do_left_v_mbuv, do_above_h_mbuv;
+    int mis = cm->mode_info_stride;
+    const MODE_INFO *mi;
+
+    // process 1st MB top-left
+    mi = mode_info_context;
+    do_left_v = (mb_col > 0);
+    do_above_h = (mb_row > 0);
+    do_left_v_mbuv = !(sb_type >= BLOCK_SIZE_SB64X64 &&
+      tx_size >= TX_32X32 && (mb_col & 2));
+    do_above_h_mbuv = !(sb_type >= BLOCK_SIZE_SB64X64 &&
+      tx_size >= TX_32X32 && (mb_row & 2));
+    lpf_mb(cm, mi, do_left_v, do_above_h,
+      do_left_v_mbuv, do_above_h_mbuv,
+      y_ptr,
+      y_only? 0 : u_ptr,
+      y_only? 0 : v_ptr,
+      y_stride, uv_stride, dering);
+    // process 2nd MB top-right
+    mi = mode_info_context + 1;
+    do_left_v = !(sb_type && (tx_size >= TX_32X32 ||
+      sb_mb_lf_skip(mode_info_context, mi)));
+    do_above_h = (mb_row > 0);
+    do_left_v_mbuv = do_left_v;
+    do_above_h_mbuv = !(sb_type >= BLOCK_SIZE_SB64X64 &&
+      tx_size >= TX_32X32 && (mb_row & 2));
+    lpf_mb(cm, mi, do_left_v, do_above_h,
+      do_left_v_mbuv, do_above_h_mbuv,
+      y_ptr + 16,
+      y_only ? 0 : (u_ptr + 8),
+      y_only ? 0 : (v_ptr + 8),
+      y_stride, uv_stride, dering);
+
+    // process 3rd MB bottom-left
+    mi = mode_info_context + mis;
+    do_left_v = (mb_col > 0);
+    do_above_h =!(sb_type && (tx_size >= TX_32X32 ||
+      sb_mb_lf_skip(mode_info_context, mi)));
+    do_left_v_mbuv =  !(sb_type >= BLOCK_SIZE_SB64X64 &&
+      tx_size >= TX_32X32 && (mb_col & 2));
+    do_above_h_mbuv = do_above_h;
+    lpf_mb(cm, mi, do_left_v, do_above_h,
+      do_left_v_mbuv, do_above_h_mbuv,
+      y_ptr + 16 * y_stride,
+      y_only ? 0 : (u_ptr + 8 * uv_stride),
+      y_only ? 0 : (v_ptr + 8 * uv_stride),
+      y_stride, uv_stride, dering);
+
+    // process 4th MB bottom right
+    mi = mode_info_context + mis + 1;
+    do_left_v = !(sb_type && (tx_size >= TX_32X32 ||
+      sb_mb_lf_skip(mi - 1, mi)));
+    do_above_h =!(sb_type && (tx_size >= TX_32X32 ||
+      sb_mb_lf_skip(mode_info_context + 1, mi)));
+    do_left_v_mbuv = do_left_v;
+    do_above_h_mbuv = do_above_h;
+    lpf_mb(cm, mi, do_left_v, do_above_h,
+      do_left_v_mbuv, do_above_h_mbuv,
+      y_ptr + 16 * y_stride + 16,
+      y_only ? 0 : (u_ptr + 8 * uv_stride + 8),
+      y_only ? 0 : (v_ptr + 8 * uv_stride + 8),
+      y_stride, uv_stride, dering);
+}
+
+static void lpf_sb64(VP9_COMMON *cm, const MODE_INFO *mode_info_context,
+                     int mb_row, int mb_col,
+                     uint8_t *y_ptr, uint8_t *u_ptr, uint8_t *v_ptr,
+                     int y_stride, int uv_stride,
+                     int y_only, int dering) {
+    lpf_sb32(cm, mode_info_context, mb_row, mb_col,
+      y_ptr, u_ptr, v_ptr,
+      y_stride, uv_stride, y_only, dering);
+    lpf_sb32(cm, mode_info_context + 2, mb_row, mb_col + 2,
+      y_ptr + 32, u_ptr + 16, v_ptr + 16,
+      y_stride, uv_stride, y_only, dering);
+    lpf_sb32(cm, mode_info_context + cm->mode_info_stride * 2,
+      mb_row + 2, mb_col,
+      y_ptr + 32 * y_stride,
+      u_ptr + 16 * uv_stride,
+      v_ptr + 16 * uv_stride,
+      y_stride, uv_stride, y_only, dering);
+    lpf_sb32(cm, mode_info_context + cm->mode_info_stride * 2 + 2,
+      mb_row + 2, mb_col + 2,
+      y_ptr + 32 * y_stride + 32,
+      u_ptr + 16 * uv_stride + 16,
+      v_ptr + 16 * uv_stride + 16,
+      y_stride, uv_stride, y_only, dering);
+}
 void vp9_loop_filter_frame(VP9_COMMON *cm,
                            MACROBLOCKD *xd,
                            int frame_filter_level,
                            int y_only,
                            int dering) {
   YV12_BUFFER_CONFIG *post = cm->frame_to_show;
-  loop_filter_info_n *lfi_n = &cm->lf_info;
-  struct loop_filter_info lfi;
-  const FRAME_TYPE frame_type = cm->frame_type;
   int mb_row, mb_col;
-
+  const int sb64_rows = cm->mb_rows / 4;
+  const int sb64_cols = cm->mb_cols / 4;
+  const int extra_sb32_row = (cm->mb_rows & 2) != 0;
+  const int extra_sb32_col = (cm->mb_cols & 2) != 0;
+  const int extra_mb_col = cm->mb_cols & 1;
+  const int extra_mb_row = cm->mb_rows & 1;
   // Set up the buffer pointers
   uint8_t *y_ptr = post->y_buffer;
   uint8_t *u_ptr = y_only ? 0 : post->u_buffer;
@@ -220,172 +408,197 @@ void vp9_loop_filter_frame(VP9_COMMON *cm,
 
   // Point at base of Mb MODE_INFO list
   const MODE_INFO *mode_info_context = cm->mi;
+  const MODE_INFO *mi;
   const int mis = cm->mode_info_stride;
+  const int y_stride = post->y_stride;
+  const int uv_stride = post->uv_stride;
+  // These two flags signal if MB left edge and above edge
+  // should be filtered using MB edge filter. Currently, MB
+  // edge filtering is not applied on MB edge internal to a
+  // 32x32 superblock if:
+  // 1) SB32 is using 32x32 prediction and 32x32 transform
+  // 2) SB32 is using 32x32 prediction and 16x16 transform
+  //    but all coefficients are zero.
+  // MB edges are on 32x32 superblock boundary are always
+  // filtered except on image frame boundary.
+  int do_left_v, do_above_h;
+  // These two flags signal if MB UV left edge and above edge
+  // should be filtered using MB edge filter. Currently, MB
+  // edge filtering is not applied for MB edges internal to
+  // a 32x32 superblock if:
+  // 1) SB32 is using 32x32 prediction and 32x32 transform
+  // 2) SB32 is using 32x32 prediction and 16x16 transform
+  //    but all coefficients are zero.
+  // 3) SB32 UV edges internal to a SB64 and 32x32 transform
+  //    is used, i.e. UV is doing 32x32 transform hence no
+  //    transform boundary exists inside the SB64 for UV
+  int do_left_v_mbuv, do_above_h_mbuv;
 
   // Initialize the loop filter for this frame.
   vp9_loop_filter_frame_init(cm, xd, frame_filter_level);
 
-  // vp9_filter each macro block
-  for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
-    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
-      const MB_PREDICTION_MODE mode = mode_info_context->mbmi.mode;
-      const int mode_index = lfi_n->mode_lf_lut[mode];
-      const int seg = mode_info_context->mbmi.segment_id;
-      const int ref_frame = mode_info_context->mbmi.ref_frame;
-      const int filter_level = lfi_n->lvl[seg][ref_frame][mode_index];
-      if (filter_level) {
-        const int skip_lf = mb_lf_skip(&mode_info_context->mbmi);
-        const int tx_size = mode_info_context->mbmi.txfm_size;
-        if (cm->filter_type == NORMAL_LOOPFILTER) {
-          const int hev_index = lfi_n->hev_thr_lut[frame_type][filter_level];
-          lfi.mblim = lfi_n->mblim[filter_level];
-          lfi.blim = lfi_n->blim[filter_level];
-          lfi.lim = lfi_n->lim[filter_level];
-          lfi.hev_thr = lfi_n->hev_thr[hev_index];
-
-          if (mb_col > 0 &&
-              !((mb_col & 1) && mode_info_context->mbmi.sb_type &&
-                (sb_mb_lf_skip(mode_info_context - 1, mode_info_context) ||
-                 tx_size >= TX_32X32))
-              ) {
-            if (tx_size >= TX_16X16)
-              vp9_lpf_mbv_w(y_ptr, u_ptr, v_ptr, post->y_stride,
-                            post->uv_stride, &lfi);
-            else
-              vp9_loop_filter_mbv(y_ptr, u_ptr, v_ptr, post->y_stride,
-                                  post->uv_stride, &lfi);
-          }
-          if (!skip_lf) {
-            if (tx_size >= TX_8X8) {
-              if (tx_size == TX_8X8 && (mode == I8X8_PRED || mode == SPLITMV))
-                vp9_loop_filter_bv8x8(y_ptr, u_ptr, v_ptr, post->y_stride,
-                                      post->uv_stride, &lfi);
-              else
-                vp9_loop_filter_bv8x8(y_ptr, NULL, NULL, post->y_stride,
-                                      post->uv_stride, &lfi);
-            } else {
-              vp9_loop_filter_bv(y_ptr, u_ptr, v_ptr, post->y_stride,
-                                 post->uv_stride, &lfi);
-            }
-          }
-          /* don't apply across umv border */
-          if (mb_row > 0 &&
-              !((mb_row & 1) && mode_info_context->mbmi.sb_type &&
-                (sb_mb_lf_skip(mode_info_context - mis, mode_info_context) ||
-                tx_size >= TX_32X32))
-              ) {
-            if (tx_size >= TX_16X16)
-              vp9_lpf_mbh_w(y_ptr, u_ptr, v_ptr, post->y_stride,
-                            post->uv_stride, &lfi);
-            else
-              vp9_loop_filter_mbh(y_ptr, u_ptr, v_ptr, post->y_stride,
-                                  post->uv_stride, &lfi);
-          }
-          if (!skip_lf) {
-            if (tx_size >= TX_8X8) {
-              if (tx_size == TX_8X8 && (mode == I8X8_PRED || mode == SPLITMV))
-                vp9_loop_filter_bh8x8(y_ptr, u_ptr, v_ptr, post->y_stride,
-                                      post->uv_stride, &lfi);
-              else
-                vp9_loop_filter_bh8x8(y_ptr, NULL, NULL, post->y_stride,
-                                      post->uv_stride, &lfi);
-            } else {
-              vp9_loop_filter_bh(y_ptr, u_ptr, v_ptr, post->y_stride,
-                                 post->uv_stride, &lfi);
-            }
-          }
-#if CONFIG_LOOP_DERING
-          if (dering) {
-            if (mb_row && mb_row < cm->mb_rows - 1 &&
-                mb_col && mb_col < cm->mb_cols - 1) {
-              vp9_post_proc_down_and_across(y_ptr, y_ptr,
-                                            post->y_stride, post->y_stride,
-                                            16, 16, dering);
-              if (!y_only) {
-                vp9_post_proc_down_and_across(u_ptr, u_ptr,
-                                              post->uv_stride, post->uv_stride,
-                                              8, 8, dering);
-                vp9_post_proc_down_and_across(v_ptr, v_ptr,
-                                              post->uv_stride, post->uv_stride,
-                                              8, 8, dering);
-              }
-            } else {
-              // Adjust the filter so that no out-of-frame data is used.
-              uint8_t *dr_y = y_ptr, *dr_u = u_ptr, *dr_v = v_ptr;
-              int w_adjust = 0;
-              int h_adjust = 0;
-
-              if (mb_col == 0) {
-                dr_y += 2;
-                dr_u += 2;
-                dr_v += 2;
-                w_adjust += 2;
-              }
-              if (mb_col == cm->mb_cols - 1)
-                w_adjust += 2;
-              if (mb_row == 0) {
-                dr_y += 2 * post->y_stride;
-                dr_u += 2 * post->uv_stride;
-                dr_v += 2 * post->uv_stride;
-                h_adjust += 2;
-              }
-              if (mb_row == cm->mb_rows - 1)
-                h_adjust += 2;
-              vp9_post_proc_down_and_across_c(dr_y, dr_y,
-                                              post->y_stride, post->y_stride,
-                                              16 - w_adjust, 16 - h_adjust,
-                                              dering);
-              if (!y_only) {
-                vp9_post_proc_down_and_across_c(dr_u, dr_u,
-                                                post->uv_stride,
-                                                post->uv_stride,
-                                                8 - w_adjust, 8 - h_adjust,
-                                                dering);
-                vp9_post_proc_down_and_across_c(dr_v, dr_v,
-                                                post->uv_stride,
-                                                post->uv_stride,
-                                                8 - w_adjust, 8 - h_adjust,
-                                                dering);
-              }
-            }
-          }
-#endif
-        } else {
-          // FIXME: Not 8x8 aware
-          if (mb_col > 0 &&
-              !(skip_lf && mb_lf_skip(&mode_info_context[-1].mbmi)) &&
-              !((mb_col & 1) && mode_info_context->mbmi.sb_type))
-            vp9_loop_filter_simple_mbv(y_ptr, post->y_stride,
-                                       lfi_n->mblim[filter_level]);
-          if (!skip_lf)
-            vp9_loop_filter_simple_bv(y_ptr, post->y_stride,
-                                      lfi_n->blim[filter_level]);
-
-          /* don't apply across umv border */
-          if (mb_row > 0 &&
-              !(skip_lf && mb_lf_skip(&mode_info_context[-mis].mbmi)) &&
-              !((mb_row & 1) && mode_info_context->mbmi.sb_type))
-            vp9_loop_filter_simple_mbh(y_ptr, post->y_stride,
-                                       lfi_n->mblim[filter_level]);
-          if (!skip_lf)
-            vp9_loop_filter_simple_bh(y_ptr, post->y_stride,
-                                      lfi_n->blim[filter_level]);
-        }
-      }
+  // vp9_filter each 64x64 SB
+  // For each SB64: the 4 SB32 are filtered in raster scan order
+  // For each SB32: the 4 MBs are filtered in raster scan order
+  // For each MB:   the left and above MB edges as well as the
+  //                internal block edges are processed together
+  for (mb_row = 0; mb_row < sb64_rows * 4; mb_row += 4) {
+    for (mb_col = 0; mb_col < sb64_cols * 4; mb_col += 4) {
+      lpf_sb64(cm, mode_info_context, mb_row, mb_col,
+               y_ptr, u_ptr, v_ptr,
+               y_stride, uv_stride, y_only, dering);
+      y_ptr += 64;
+      u_ptr = y_only? 0 : u_ptr + 32;
+      v_ptr = y_only? 0 : v_ptr + 32;
+      mode_info_context += 4;       // step to next SB64
+    }
+    if (extra_sb32_col) {
+      // process 2 SB32s in the extra SB32 col
+      lpf_sb32(cm, mode_info_context, mb_row, mb_col,
+               y_ptr, u_ptr, v_ptr,
+               y_stride, uv_stride, y_only, dering);
+      lpf_sb32(cm, mode_info_context + mis * 2,
+               mb_row + 2, mb_col,
+               y_ptr + 32 * y_stride,
+               u_ptr + 16 * uv_stride,
+               v_ptr + 16 * uv_stride,
+               y_stride, uv_stride, y_only, dering);
+      y_ptr += 32;
+      u_ptr = y_only? 0 : u_ptr + 16;
+      v_ptr = y_only? 0 : v_ptr + 16;
+      mode_info_context += 2;       // step to next SB32
+      mb_col += 2;
+    }
+    if (extra_mb_col) {
+      // process 4 MB in the extra MB col
+      // process 1st MB
+      mi = mode_info_context;
+      do_left_v = (mb_col > 0);
+      do_above_h = (mb_row > 0);
+      do_left_v_mbuv =  1;
+      do_above_h_mbuv = 1;
+      lpf_mb(cm, mi, do_left_v, do_above_h,
+             do_left_v_mbuv, do_above_h_mbuv,
+             y_ptr,
+             y_only? 0 : u_ptr,
+             y_only? 0 : v_ptr,
+             y_stride, uv_stride, dering);
+      // process 2nd MB
+      mi = mode_info_context + mis;
+      do_left_v = (mb_col > 0);
+      do_above_h = 1;
+      do_left_v_mbuv =  1;
+      do_above_h_mbuv = 1;
+      lpf_mb(cm, mi, do_left_v, do_above_h,
+             do_left_v_mbuv, do_above_h_mbuv,
+             y_ptr + 16 * y_stride,
+             y_only ? 0 : (u_ptr + 8 * uv_stride),
+             y_only ? 0 : (v_ptr + 8 * uv_stride),
+             y_stride, uv_stride, dering);
+      // process 3nd MB
+      mi = mode_info_context + mis * 2;
+      do_left_v = (mb_col > 0);
+      do_above_h = 1;
+      do_left_v_mbuv =  1;
+      do_above_h_mbuv = 1;
+      lpf_mb(cm, mi, do_left_v, do_above_h,
+             do_left_v_mbuv, do_above_h_mbuv,
+             y_ptr + 32 * y_stride,
+             y_only ? 0 : (u_ptr + 16 * uv_stride),
+             y_only ? 0 : (v_ptr + 16 * uv_stride),
+             y_stride, uv_stride, dering);
+      // process 4th MB
+      mi = mode_info_context + mis * 3;
+      do_left_v = (mb_col > 0);
+      do_above_h = 1;
+      do_left_v_mbuv =  1;
+      do_above_h_mbuv = 1;
+      lpf_mb(cm, mi, do_left_v, do_above_h,
+             do_left_v_mbuv, do_above_h_mbuv,
+             y_ptr + 48 * y_stride,
+             y_only ? 0 : (u_ptr + 24 * uv_stride),
+             y_only ? 0 : (v_ptr + 24 * uv_stride),
+             y_stride, uv_stride, dering);
       y_ptr += 16;
-      if (!y_only) {
-        u_ptr += 8;
-        v_ptr += 8;
-      }
-      mode_info_context++;  // step to next MB
+      u_ptr = y_only? 0 : u_ptr + 8;
+      v_ptr = y_only? 0 : v_ptr + 8;
+      mode_info_context++;       // step to next MB
     }
-    y_ptr += post->y_stride  * 16 - post->y_width;
+    // move pointers to the begining of next sb64 row
+    y_ptr += y_stride  * 64 - post->y_width;
     if (!y_only) {
-      u_ptr += post->uv_stride *  8 - post->uv_width;
-      v_ptr += post->uv_stride *  8 - post->uv_width;
+      u_ptr += uv_stride *  32 - post->uv_width;
+      v_ptr += uv_stride *  32 - post->uv_width;
+    }
+    /* skip to next SB64 row */
+    mode_info_context += mis * 4 - cm->mb_cols;
+  }
+  if (extra_sb32_row) {
+    const int sb32_cols = sb64_cols * 2 + extra_sb32_col;
+    for (mb_col = 0; mb_col < sb32_cols * 2; mb_col += 2) {
+      lpf_sb32(cm, mode_info_context, mb_row, mb_col,
+               y_ptr, u_ptr, v_ptr,
+               y_stride, uv_stride, y_only, dering);
+      y_ptr += 32;
+      u_ptr = y_only? 0 : u_ptr + 16;
+      v_ptr = y_only? 0 : v_ptr + 16;
+      mode_info_context += 2;       // step to next SB32
+    }
+    if (extra_mb_col) {
+      // process 1st MB
+      mi = mode_info_context;
+      do_left_v = (mb_col > 0);
+      do_above_h = (mb_row > 0);
+      do_left_v_mbuv =  1;
+      do_above_h_mbuv = 1;
+      lpf_mb(cm, mi, do_left_v, do_above_h,
+             do_left_v_mbuv, do_above_h_mbuv,
+             y_ptr,
+             y_only? NULL : u_ptr,
+             y_only? NULL : v_ptr,
+             y_stride, uv_stride, dering);
+      // process 2nd MB
+      mi = mode_info_context + mis;
+      do_left_v = (mb_col > 0);
+      do_above_h = 1;
+      do_left_v_mbuv =  1;
+      do_above_h_mbuv = 1;
+      lpf_mb(cm, mi, do_left_v, do_above_h,
+             do_left_v_mbuv, do_above_h_mbuv,
+             y_ptr + 16 * y_stride,
+             y_only ? NULL : (u_ptr + 8 * uv_stride),
+             y_only ? NULL : (v_ptr + 8 * uv_stride),
+             y_stride, uv_stride, dering);
+      y_ptr += 16;
+      u_ptr = y_only? 0 : u_ptr + 8;
+      v_ptr = y_only? 0 : v_ptr + 8;
+      mode_info_context++;       /* step to next MB */
+    }
+    // move pointers to the beginning of next sb64 row
+    y_ptr += y_stride * 32 - post->y_width;
+    u_ptr += y_only? 0 : uv_stride *  16 - post->uv_width;
+    v_ptr += y_only? 0 : uv_stride *  16 - post->uv_width;
+    // skip to next MB row if exist
+    mode_info_context += mis * 2 - cm->mb_cols;
+    mb_row += 2;
+  }
+  if (extra_mb_row) {
+    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
+      const MODE_INFO *mi = mode_info_context;
+      do_left_v =  (mb_col > 0);
+      do_above_h = (mb_row > 0);
+      do_left_v_mbuv = 1;
+      do_above_h_mbuv = 1;
+      lpf_mb(cm, mi, do_left_v, do_above_h,
+             do_left_v_mbuv, do_above_h_mbuv,
+             y_ptr,
+             y_only? 0 : u_ptr,
+             y_only? 0 : v_ptr,
+             y_stride, uv_stride, dering);
+      y_ptr += 16;
+      u_ptr = y_only? 0 : u_ptr + 8;
+      v_ptr = y_only? 0 : v_ptr + 8;
+      mode_info_context++;     // step to next MB
     }
-    mode_info_context++;  // Skip border mb
   }
 }
-
-