shithub: libvpx

--- a/vp9/encoder/vp9_encodeframe.c

+++ b/vp9/encoder/vp9_encodeframe.c

@@ -3318,7 +3318,10 @@

       vp9_setup_pre_planes(xd, ref, cfg, mi_row, mi_col,

                            &xd->block_refs[ref]->sf);

-    vp9_build_inter_predictors_sb(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));

+    if (!cpi->sf.reuse_inter_pred_sby)

+      vp9_build_inter_predictors_sby(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));

+    vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));

     if (!x->skip) {

       mbmi->skip = 1;

--- a/vp9/encoder/vp9_pickmode.c

+++ b/vp9/encoder/vp9_pickmode.c

@@ -23,6 +23,7 @@

 #include "vp9/common/vp9_reconintra.h"

 #include "vp9/encoder/vp9_encoder.h"

+#include "vp9/encoder/vp9_pickmode.h"

 #include "vp9/encoder/vp9_ratectrl.h"

 #include "vp9/encoder/vp9_rdopt.h"

@@ -183,6 +184,22 @@

   *out_dist_sum += dist << 4;

+static int get_pred_buffer(PRED_BUFFER *p, int len) {

+  int i;

+  for (i = 0; i < len; i++) {

+    if (!p[i].in_use) {

+      p[i].in_use = 1;

+      return i;

+    }

+  }

+  return -1;

+}

+static void free_pred_buffer(PRED_BUFFER *p) {

+  p->in_use = 0;

+}

 // TODO(jingning) placeholder for inter-frame non-RD mode decision.

 // this needs various further optimizations. to be continued..

 int64_t vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,

@@ -229,6 +246,31 @@

   const int pred_filter_search = (((mi_row + mi_col) >> bsl) +

                                       get_chessboard_index(cm)) % 2;

+  // For speed 6, the result of interp filter is reused later in actual encoding

+  // process.

+  int bh = num_4x4_blocks_high_lookup[bsize] << 2;

+  int bw = num_4x4_blocks_wide_lookup[bsize] << 2;

+  int pixels_in_block = bh * bw;

+  // tmp[3] points to dst buffer, and the other 3 point to allocated buffers.

+  PRED_BUFFER tmp[4];

+  DECLARE_ALIGNED_ARRAY(16, uint8_t, pred_buf, 3 * 64 * 64);

+  struct buf_2d orig_dst = pd->dst;

+  PRED_BUFFER *best_pred = NULL;

+  PRED_BUFFER *this_mode_pred = NULL;

+  int i;

+  if (cpi->sf.reuse_inter_pred_sby) {

+    for (i = 0; i < 3; i++) {

+      tmp[i].data = &pred_buf[pixels_in_block * i];

+      tmp[i].stride = bw;

+      tmp[i].in_use = 0;

+    }

+    tmp[3].data = pd->dst.buf;

+    tmp[3].stride = pd->dst.stride;

+    tmp[3].in_use = 0;

+  }

   x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;

   x->skip = 0;

@@ -324,6 +366,16 @@

       // Search for the best prediction filter type, when the resulting

       // motion vector is at sub-pixel accuracy level for luma component, i.e.,

       // the last three bits are all zeros.

+      if (cpi->sf.reuse_inter_pred_sby) {

+        if (this_mode == NEARESTMV) {

+          this_mode_pred = &tmp[3];

+        } else {

+          this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];

+          pd->dst.buf = this_mode_pred->data;

+          pd->dst.stride = bw;

+        }

+      }

       if ((this_mode == NEWMV || filter_ref == SWITCHABLE) &&

           pred_filter_search &&

           ((mbmi->mv[0].as_mv.row & 0x07) != 0 ||

@@ -334,6 +386,7 @@

         unsigned int pf_sse[3];

         int64_t best_cost = INT64_MAX;

         INTERP_FILTER best_filter = SWITCHABLE, filter;

+        PRED_BUFFER *current_pred = this_mode_pred;

         for (filter = EIGHTTAP; filter <= EIGHTTAP_SHARP; ++filter) {

           int64_t cost;

@@ -345,12 +398,28 @@

                         vp9_get_switchable_rate(cpi) + pf_rate[filter],

                         pf_dist[filter]);

           if (cost < best_cost) {

-              best_filter = filter;

-              best_cost = cost;

-              skip_txfm = x->skip_txfm;

+            best_filter = filter;

+            best_cost = cost;

+            skip_txfm = x->skip_txfm;

+            if (cpi->sf.reuse_inter_pred_sby) {

+              if (this_mode_pred != current_pred) {

+                free_pred_buffer(this_mode_pred);

+                this_mode_pred = current_pred;

+              }

+              if (filter < EIGHTTAP_SHARP) {

+                current_pred = &tmp[get_pred_buffer(tmp, 3)];

+                pd->dst.buf = current_pred->data;

+                pd->dst.stride = bw;

+              }

+            }

+        if (cpi->sf.reuse_inter_pred_sby && this_mode_pred != current_pred)

+          free_pred_buffer(current_pred);

         mbmi->interp_filter = best_filter;

         rate = pf_rate[mbmi->interp_filter];

         dist = pf_dist[mbmi->interp_filter];

@@ -449,6 +518,16 @@

         best_pred_filter = mbmi->interp_filter;

         best_ref_frame = ref_frame;

         skip_txfm = x->skip_txfm;

+        if (cpi->sf.reuse_inter_pred_sby) {

+          if (best_pred != NULL)

+            free_pred_buffer(best_pred);

+          best_pred = this_mode_pred;

+        }

+      } else {

+        if (cpi->sf.reuse_inter_pred_sby)

+          free_pred_buffer(this_mode_pred);

       if (x->skip)

@@ -456,7 +535,20 @@

+  // If best prediction is not in dst buf, then copy the prediction block from

+  // temp buf to dst buf.

+  if (cpi->sf.reuse_inter_pred_sby && best_pred->data != orig_dst.buf) {

+    uint8_t *copy_from, *copy_to;

+    pd->dst = orig_dst;

+    copy_to = pd->dst.buf;

+    copy_from = best_pred->data;

+    vp9_convolve_copy(copy_from, bw, copy_to, pd->dst.stride, NULL, 0, NULL, 0,

+                      bw, bh);

+  }

   mbmi->mode = best_mode;

   mbmi->interp_filter = best_pred_filter;

   mbmi->ref_frame[0] = best_ref_frame;

@@ -469,6 +561,11 @@

   if (!x->skip && best_rd > inter_mode_thresh &&

       bsize <= cpi->sf.max_intra_bsize) {

     for (this_mode = DC_PRED; this_mode <= DC_PRED; ++this_mode) {

+      if (cpi->sf.reuse_inter_pred_sby) {

+        pd->dst.buf = tmp[0].data;

+        pd->dst.stride = bw;

+      }

       vp9_predict_intra_block(xd, 0, b_width_log2(bsize),

                               mbmi->tx_size, this_mode,

                               &p->src.buf[0], p->src.stride,

@@ -475,6 +572,10 @@

                               &pd->dst.buf[0], pd->dst.stride, 0, 0, 0);

       model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist, &var_y, &sse_y);

+      if (cpi->sf.reuse_inter_pred_sby)

+        pd->dst = orig_dst;

       rate += cpi->mbmode_cost[this_mode];

       rate += intra_cost_penalty;

       this_rd = RDCOST(x->rdmult, x->rddiv, rate, dist);

@@ -492,6 +593,7 @@

 #if CONFIG_DENOISING

   vp9_denoiser_denoise(&cpi->denoiser, x, mi_row, mi_col, bsize);

 #endif

--- a/vp9/encoder/vp9_pickmode.h

+++ b/vp9/encoder/vp9_pickmode.h

@@ -17,6 +17,12 @@

 extern "C" {

 #endif

+typedef struct {

+  uint8_t *data;

+  int stride;

+  int in_use;

+} PRED_BUFFER;

 int64_t vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,

                             const struct TileInfo *const tile,

                             int mi_row, int mi_col,

--- a/vp9/encoder/vp9_speed_features.c

+++ b/vp9/encoder/vp9_speed_features.c

@@ -277,6 +277,9 @@

     // is checked for a partition block. Later, we can try to allow large

     // partitions to do intra mode checking.

     sf->max_intra_bsize = BLOCK_8X8;

+    // This feature is only enabled when partition search is disabled.

+    sf->reuse_inter_pred_sby = 1;

   if (speed >= 7) {

@@ -342,6 +345,7 @@

   for (i = 0; i < BLOCK_SIZES; ++i)

     sf->inter_mode_mask[i] = INTER_ALL;

   sf->max_intra_bsize = BLOCK_64X64;

+  sf->reuse_inter_pred_sby = 0;

   // This setting only takes effect when partition_search_type is set

   // to FIXED_PARTITION.

   sf->always_this_block_size = BLOCK_16X16;

--- a/vp9/encoder/vp9_speed_features.h

+++ b/vp9/encoder/vp9_speed_features.h

@@ -353,6 +353,11 @@

   // The threshold used in SOURCE_VAR_BASED_PARTITION search type.

   unsigned int source_var_thresh;

+  // When partition is pre-set, the inter prediction result from pick_inter_mode

+  // can be reused in final block encoding process. It is enabled only for real-

+  // time mode speed 6.

+  int reuse_inter_pred_sby;

 } SPEED_FEATURES;

 struct VP9_COMP;

--

⑨