diff --git a/vp9/encoder/vp9_block.h b/vp9/encoder/vp9_block.h
index b88506b3b..98662c94a 100644
--- a/vp9/encoder/vp9_block.h
+++ b/vp9/encoder/vp9_block.h
@@ -32,12 +32,13 @@ typedef struct {
   int16_t *dqcoeff[MAX_MB_PLANE][2];
   uint16_t *eobs[MAX_MB_PLANE][2];
 
-  // dual buffer pointers
+  // dual buffer pointers, 0: in use, 1: best in store
   int16_t *coeff_pbuf[MAX_MB_PLANE][2];
   int16_t *qcoeff_pbuf[MAX_MB_PLANE][2];
   int16_t *dqcoeff_pbuf[MAX_MB_PLANE][2];
   uint16_t *eobs_pbuf[MAX_MB_PLANE][2];
 
+  int is_coded;
   int num_4x4_blk;
   int skip;
   int_mv best_ref_mv;
@@ -92,6 +93,8 @@ struct macroblock {
 
   MACROBLOCKD e_mbd;
   int skip_block;
+  int select_txfm_size;
+  int skip_optimize;
 
   search_site *ss;
   int ss_count;
diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c
index 2e2ecf8cd..9d806508e 100644
--- a/vp9/encoder/vp9_encodeframe.c
+++ b/vp9/encoder/vp9_encodeframe.c
@@ -613,11 +613,12 @@ static void pick_sb_modes(VP9_COMP *cpi, const TileInfo *const tile,
   xd->mi_8x8[0]->mbmi.sb_type = bsize;
 
   for (i = 0; i < MAX_MB_PLANE; ++i) {
-    p[i].coeff = ctx->coeff_pbuf[i][1];
-    pd[i].qcoeff = ctx->qcoeff_pbuf[i][1];
-    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
-    pd[i].eobs = ctx->eobs_pbuf[i][1];
+    p[i].coeff = ctx->coeff_pbuf[i][0];
+    pd[i].qcoeff = ctx->qcoeff_pbuf[i][0];
+    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
+    pd[i].eobs = ctx->eobs_pbuf[i][0];
   }
+  ctx->is_coded = 0;
 
   // Set to zero to make sure we do not use the previous encoded frame stats
   xd->mi_8x8[0]->mbmi.skip_coeff = 0;
@@ -2400,10 +2401,13 @@ static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, int output_enabled,
   MODE_INFO **mi_8x8 = xd->mi_8x8;
   MODE_INFO *mi = mi_8x8[0];
   MB_MODE_INFO *mbmi = &mi->mbmi;
+  PICK_MODE_CONTEXT *ctx = get_block_context(x, bsize);
   unsigned int segment_id = mbmi->segment_id;
   const int mis = cm->mode_info_stride;
   const int mi_width = num_8x8_blocks_wide_lookup[bsize];
   const int mi_height = num_8x8_blocks_high_lookup[bsize];
+  x->skip_optimize = ctx->is_coded;
+  ctx->is_coded = 1;
   x->use_lp32x32fdct = cpi->sf.use_lp32x32fdct;
   x->skip_encode = (!output_enabled && cpi->sf.skip_encode_frame &&
                     xd->q_index < QIDX_SKIP_THRESH);
diff --git a/vp9/encoder/vp9_encodemb.c b/vp9/encoder/vp9_encodemb.c
index 70008103e..7afed26d6 100644
--- a/vp9/encoder/vp9_encodemb.c
+++ b/vp9/encoder/vp9_encodemb.c
@@ -432,18 +432,26 @@ static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize,
   // TODO(jingning): per transformed block zero forcing only enabled for
   // luma component. will integrate chroma components as well.
   if (x->zcoeff_blk[tx_size][block] && plane == 0) {
-    int i, j;
+    int i, k;
     pd->eobs[block] = 0;
-    txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
+    txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &k);
     ctx->ta[plane][i] = 0;
-    ctx->tl[plane][j] = 0;
+    ctx->tl[plane][k] = 0;
     return;
   }
 
-  vp9_xform_quant(plane, block, plane_bsize, tx_size, arg);
+  if (x->select_txfm_size || xd->mi_8x8[0]->mbmi.sb_type < BLOCK_8X8)
+    vp9_xform_quant(plane, block, plane_bsize, tx_size, arg);
 
-  if (x->optimize)
+  if (x->optimize && (x->select_txfm_size ||
+      xd->mi_8x8[0]->mbmi.sb_type < BLOCK_8X8|| !x->skip_optimize)) {
     vp9_optimize_b(plane, block, plane_bsize, tx_size, x, ctx);
+  } else {
+    int i, k;
+    txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &k);
+    ctx->ta[plane][i] = pd->eobs[block] > 0;
+    ctx->tl[plane][k] = pd->eobs[block] > 0;
+  }
 
   if (x->skip_encode || pd->eobs[block] == 0)
     return;
@@ -507,7 +515,8 @@ void vp9_encode_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
   struct optimize_ctx ctx;
   struct encode_b_args arg = {x, &ctx};
 
-  vp9_subtract_sb(x, bsize);
+  if (x->select_txfm_size || xd->mi_8x8[0]->mbmi.sb_type < BLOCK_8X8)
+    vp9_subtract_sb(x, bsize);
 
   if (x->optimize) {
     int i;
diff --git a/vp9/encoder/vp9_onyx_if.c b/vp9/encoder/vp9_onyx_if.c
index 651ce5ed8..8a9834310 100644
--- a/vp9/encoder/vp9_onyx_if.c
+++ b/vp9/encoder/vp9_onyx_if.c
@@ -834,6 +834,7 @@ void vp9_set_speed_features(VP9_COMP *cpi) {
 
         sf->adaptive_rd_thresh = 2;
         sf->recode_loop = 2;
+        sf->use_lp32x32fdct = 1;
         sf->mode_skip_start = 11;
         sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V;
         sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V;
diff --git a/vp9/encoder/vp9_rdopt.c b/vp9/encoder/vp9_rdopt.c
index 2155a9c11..3565c0762 100644
--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -246,6 +246,9 @@ void vp9_initialize_rd_consts(VP9_COMP *cpi) {
 
   vp9_set_speed_features(cpi);
 
+  cpi->mb.select_txfm_size = cpi->sf.tx_size_search_method == USE_LARGESTALL ?
+                             0 : 1;
+
   set_block_thresholds(cpi);
 
   fill_token_costs(cpi->mb.token_costs, cm->fc.coef_probs);
@@ -3030,6 +3033,29 @@ static int64_t handle_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
   return this_rd;  // if 0, this will be re-calculated by caller
 }
 
+static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
+  int i;
+  struct macroblock_plane *const p = x->plane;
+  struct macroblockd_plane *const pd = x->e_mbd.plane;
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    p[i].coeff    = ctx->coeff_pbuf[i][1];
+    pd[i].qcoeff  = ctx->qcoeff_pbuf[i][1];
+    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
+    pd[i].eobs    = ctx->eobs_pbuf[i][1];
+
+    ctx->coeff_pbuf[i][1]   = ctx->coeff_pbuf[i][0];
+    ctx->qcoeff_pbuf[i][1]  = ctx->qcoeff_pbuf[i][0];
+    ctx->dqcoeff_pbuf[i][1] = ctx->dqcoeff_pbuf[i][0];
+    ctx->eobs_pbuf[i][1]    = ctx->eobs_pbuf[i][0];
+
+    ctx->coeff_pbuf[i][0]   = p[i].coeff;
+    ctx->qcoeff_pbuf[i][0]  = pd[i].qcoeff;
+    ctx->dqcoeff_pbuf[i][0] = pd[i].dqcoeff;
+    ctx->eobs_pbuf[i][0]    = pd[i].eobs;
+  }
+}
+
 void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
                                int *returnrate, int64_t *returndist,
                                BLOCK_SIZE bsize,
@@ -3572,6 +3598,8 @@ int64_t vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, MACROBLOCK *x,
         best_rd = this_rd;
         best_mbmode = *mbmi;
         best_skip2 = this_skip2;
+        if (!x->select_txfm_size)
+          swap_block_ptr(x, ctx);
         vpx_memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mbmi->tx_size],
                    sizeof(uint8_t) * ctx->num_4x4_blk);
 
@@ -4316,6 +4344,8 @@ int64_t vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
                    RDCOST(x->rdmult, x->rddiv, rate_uv, distortion_uv);
         best_mbmode = *mbmi;
         best_skip2 = this_skip2;
+        if (!x->select_txfm_size)
+          swap_block_ptr(x, ctx);
         vpx_memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mbmi->tx_size],
                    sizeof(uint8_t) * ctx->num_4x4_blk);