Moving txfm_size bits before prediction mode bits.

Hardware implementation needs to load coeff probs based on the
transform size. For selectable transform size, moving these bits
earlier in the bitstream adds some delay giving time to preload
the probs and speeds up the decoding process.

Change-Id: I3bfc1f662ae6f219c9286fe9ae6310c7d8a63ea7

diff --git a/vp9/decoder/vp9_decodemv.c b/vp9/decoder/vp9_decodemv.c
index 2bf30866970bdff36a0ad2afb7528f5ad55a84a1..9f6a0f2cc847709bb1ac4742a90db0209c7aaf42 100644 (file)
--- a/vp9/decoder/vp9_decodemv.c
+++ b/vp9/decoder/vp9_decodemv.c
@@ -118,6 +118,24 @@ static void kfread_modes(VP9D_COMP *pbi, MODE_INFO *m,
   if (!m->mbmi.mb_skip_coeff)
     m->mbmi.mb_skip_coeff = vp9_read(r, vp9_get_pred_prob(cm, xd, PRED_MBSKIP));
 
+  if (cm->txfm_mode == TX_MODE_SELECT &&
+      m->mbmi.sb_type >= BLOCK_SIZE_SB8X8) {
+    const int allow_16x16 = m->mbmi.sb_type >= BLOCK_SIZE_MB16X16;
+    const int allow_32x32 = m->mbmi.sb_type >= BLOCK_SIZE_SB32X32;
+    m->mbmi.txfm_size = select_txfm_size(cm, r, allow_16x16, allow_32x32);
+  } else if (cm->txfm_mode >= ALLOW_32X32 &&
+             m->mbmi.sb_type >= BLOCK_SIZE_SB32X32) {
+    m->mbmi.txfm_size = TX_32X32;
+  } else if (cm->txfm_mode >= ALLOW_16X16 &&
+             m->mbmi.sb_type >= BLOCK_SIZE_MB16X16) {
+    m->mbmi.txfm_size = TX_16X16;
+  } else if (cm->txfm_mode >= ALLOW_8X8 &&
+             m->mbmi.sb_type >= BLOCK_SIZE_SB8X8) {
+    m->mbmi.txfm_size = TX_8X8;
+  } else {
+    m->mbmi.txfm_size = TX_4X4;
+  }
+
   // luma mode
   if (m->mbmi.sb_type >= BLOCK_SIZE_SB8X8) {
     const MB_PREDICTION_MODE A = above_block_mode(m, 0, mis);
@@ -153,25 +171,6 @@ static void kfread_modes(VP9D_COMP *pbi, MODE_INFO *m,
   }
 
   m->mbmi.uv_mode = read_uv_mode(r, cm->kf_uv_mode_prob[m->mbmi.mode]);
-
-  if (cm->txfm_mode == TX_MODE_SELECT &&
-      m->mbmi.sb_type >= BLOCK_SIZE_SB8X8) {
-    const int allow_16x16 = m->mbmi.sb_type >= BLOCK_SIZE_MB16X16;
-    const int allow_32x32 = m->mbmi.sb_type >= BLOCK_SIZE_SB32X32;
-    m->mbmi.txfm_size = select_txfm_size(cm, r, allow_16x16, allow_32x32);
-  } else if (cm->txfm_mode >= ALLOW_32X32 &&
-             m->mbmi.sb_type >= BLOCK_SIZE_SB32X32) {
-    m->mbmi.txfm_size = TX_32X32;
-  } else if (cm->txfm_mode >= ALLOW_16X16 &&
-             m->mbmi.sb_type >= BLOCK_SIZE_MB16X16 &&
-             m->mbmi.mode <= TM_PRED) {
-    m->mbmi.txfm_size = TX_16X16;
-  } else if (cm->txfm_mode >= ALLOW_8X8 &&
-             m->mbmi.sb_type >= BLOCK_SIZE_SB8X8) {
-    m->mbmi.txfm_size = TX_8X8;
-  } else {
-    m->mbmi.txfm_size = TX_4X4;
-  }
 }
 
 static int read_mv_component(vp9_reader *r,
@@ -569,6 +568,24 @@ static void read_mb_modes_mv(VP9D_COMP *pbi, MODE_INFO *mi, MB_MODE_INFO *mbmi,
   // Read the reference frame
   mbmi->ref_frame = read_ref_frame(pbi, r, mbmi->segment_id);
 
+  if (cm->txfm_mode == TX_MODE_SELECT &&
+      (mbmi->mb_skip_coeff == 0 || mbmi->ref_frame == INTRA_FRAME) &&
+      bsize >= BLOCK_SIZE_SB8X8) {
+    const int allow_16x16 = bsize >= BLOCK_SIZE_MB16X16;
+    const int allow_32x32 = bsize >= BLOCK_SIZE_SB32X32;
+    mbmi->txfm_size = select_txfm_size(cm, r, allow_16x16, allow_32x32);
+  } else if (bsize >= BLOCK_SIZE_SB32X32 &&
+             cm->txfm_mode >= ALLOW_32X32) {
+    mbmi->txfm_size = TX_32X32;
+  } else if (cm->txfm_mode >= ALLOW_16X16 &&
+             bsize >= BLOCK_SIZE_MB16X16) {
+    mbmi->txfm_size = TX_16X16;
+  } else if (cm->txfm_mode >= ALLOW_8X8 && (bsize >= BLOCK_SIZE_SB8X8)) {
+    mbmi->txfm_size = TX_8X8;
+  } else {
+    mbmi->txfm_size = TX_4X4;
+  }
+
   // If reference frame is an Inter frame
   if (mbmi->ref_frame) {
     int_mv nearest, nearby, best_mv;
@@ -833,24 +850,6 @@ static void read_mb_modes_mv(VP9D_COMP *pbi, MODE_INFO *mi, MB_MODE_INFO *mbmi,
     mbmi->uv_mode = read_uv_mode(r, cm->fc.uv_mode_prob[mbmi->mode]);
     cm->fc.uv_mode_counts[mbmi->mode][mbmi->uv_mode]++;
   }
-
-  if (cm->txfm_mode == TX_MODE_SELECT &&
-      (mbmi->mb_skip_coeff == 0 || mbmi->ref_frame == INTRA_FRAME) &&
-      bsize >= BLOCK_SIZE_SB8X8) {
-    const int allow_16x16 = bsize >= BLOCK_SIZE_MB16X16;
-    const int allow_32x32 = bsize >= BLOCK_SIZE_SB32X32;
-    mbmi->txfm_size = select_txfm_size(cm, r, allow_16x16, allow_32x32);
-  } else if (bsize >= BLOCK_SIZE_SB32X32 &&
-             cm->txfm_mode >= ALLOW_32X32) {
-    mbmi->txfm_size = TX_32X32;
-  } else if (cm->txfm_mode >= ALLOW_16X16 &&
-             bsize >= BLOCK_SIZE_MB16X16) {
-    mbmi->txfm_size = TX_16X16;
-  } else if (cm->txfm_mode >= ALLOW_8X8 && (bsize >= BLOCK_SIZE_SB8X8)) {
-    mbmi->txfm_size = TX_8X8;
-  } else {
-    mbmi->txfm_size = TX_4X4;
-  }
 }
 
 void vp9_decode_mode_mvs_init(VP9D_COMP* const pbi, vp9_reader *r) {

diff --git a/vp9/encoder/vp9_bitstream.c b/vp9/encoder/vp9_bitstream.c
index 49b2d29eca7023c75c824b5032fd7a2da3777ddd..48f70e2e42a4d0cc10bbcbdc3fa97446456a19ba 100644 (file)
--- a/vp9/encoder/vp9_bitstream.c
+++ b/vp9/encoder/vp9_bitstream.c
@@ -700,6 +700,19 @@ static void pack_inter_mode_mvs(VP9_COMP *cpi, MODE_INFO *m,
   // Encode the reference frame.
   encode_ref_frame(bc, pc, xd, segment_id, rf);
 
+  if (mi->sb_type >= BLOCK_SIZE_SB8X8 && pc->txfm_mode == TX_MODE_SELECT &&
+      !(rf != INTRA_FRAME &&
+        (skip_coeff || vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP)))) {
+    TX_SIZE sz = mi->txfm_size;
+    // FIXME(rbultje) code ternary symbol once all experiments are merged
+    vp9_write(bc, sz != TX_4X4, pc->prob_tx[0]);
+    if (mi->sb_type >= BLOCK_SIZE_MB16X16 && sz != TX_4X4) {
+      vp9_write(bc, sz != TX_8X8, pc->prob_tx[1]);
+      if (mi->sb_type >= BLOCK_SIZE_SB32X32 && sz != TX_8X8)
+        vp9_write(bc, sz != TX_16X16, pc->prob_tx[2]);
+    }
+  }
+
   if (rf == INTRA_FRAME) {
 #ifdef ENTROPY_STATS
     active_section = 6;
@@ -807,19 +820,6 @@ static void pack_inter_mode_mvs(VP9_COMP *cpi, MODE_INFO *m,
         break;
     }
   }
-
-  if (mi->sb_type >= BLOCK_SIZE_SB8X8 && pc->txfm_mode == TX_MODE_SELECT &&
-      !(rf != INTRA_FRAME &&
-        (skip_coeff || vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP)))) {
-    TX_SIZE sz = mi->txfm_size;
-    // FIXME(rbultje) code ternary symbol once all experiments are merged
-    vp9_write(bc, sz != TX_4X4, pc->prob_tx[0]);
-    if (mi->sb_type >= BLOCK_SIZE_MB16X16 && sz != TX_4X4) {
-      vp9_write(bc, sz != TX_8X8, pc->prob_tx[1]);
-      if (mi->sb_type >= BLOCK_SIZE_SB32X32 && sz != TX_8X8)
-        vp9_write(bc, sz != TX_16X16, pc->prob_tx[2]);
-    }
-  }
 }
 
 static void write_mb_modes_kf(const VP9_COMP *cpi,
@@ -842,6 +842,17 @@ static void write_mb_modes_kf(const VP9_COMP *cpi,
     vp9_write(bc, skip_coeff, vp9_get_pred_prob(c, xd, PRED_MBSKIP));
   }
 
+  if (m->mbmi.sb_type >= BLOCK_SIZE_SB8X8 && c->txfm_mode == TX_MODE_SELECT) {
+    TX_SIZE sz = m->mbmi.txfm_size;
+    // FIXME(rbultje) code ternary symbol once all experiments are merged
+    vp9_write(bc, sz != TX_4X4, c->prob_tx[0]);
+    if (m->mbmi.sb_type >= BLOCK_SIZE_MB16X16 && sz != TX_4X4) {
+      vp9_write(bc, sz != TX_8X8, c->prob_tx[1]);
+      if (m->mbmi.sb_type >= BLOCK_SIZE_SB32X32 && sz != TX_8X8)
+        vp9_write(bc, sz != TX_16X16, c->prob_tx[2]);
+    }
+  }
+
   if (m->mbmi.sb_type >= BLOCK_SIZE_SB8X8) {
     const B_PREDICTION_MODE A = above_block_mode(m, 0, mis);
     const B_PREDICTION_MODE L = xd->left_available ?
@@ -866,17 +877,6 @@ static void write_mb_modes_kf(const VP9_COMP *cpi,
   }
 
   write_uv_mode(bc, m->mbmi.uv_mode, c->kf_uv_mode_prob[ym]);
-
-  if (m->mbmi.sb_type >= BLOCK_SIZE_SB8X8 && c->txfm_mode == TX_MODE_SELECT) {
-    TX_SIZE sz = m->mbmi.txfm_size;
-    // FIXME(rbultje) code ternary symbol once all experiments are merged
-    vp9_write(bc, sz != TX_4X4, c->prob_tx[0]);
-    if (m->mbmi.sb_type >= BLOCK_SIZE_MB16X16 && sz != TX_4X4) {
-      vp9_write(bc, sz != TX_8X8, c->prob_tx[1]);
-      if (m->mbmi.sb_type >= BLOCK_SIZE_SB32X32 && sz != TX_8X8)
-        vp9_write(bc, sz != TX_16X16, c->prob_tx[2]);
-    }
-  }
 }
 
 static void write_modes_b(VP9_COMP *cpi, MODE_INFO *m, vp9_writer *bc,