shithub: libvpx

Info  •  Files  •  Log  •  Branches

SVC 1 pass mode: Constrain inter mode search within superframe.

Keep track of frame indexes for the references, and
constrain inter mode search for reference with same
temporal alignment.

Improves speed by about ~15%, no noticeable loss in
compression performance.

Change-Id: I5c407a8acca921234060c4fcef4afd7d734201c8

--- a/vp9/encoder/vp9_encoder.c

+++ b/vp9/encoder/vp9_encoder.c

@@ -2793,6 +2793,22 @@

                                    cpi->resize_pending);

   }

 #endif

+  if (is_one_pass_cbr_svc(cpi)) {

+    // Keep track of frame index for each reference frame.

+    SVC *const svc = &cpi->svc;

+    if (cm->frame_type == KEY_FRAME) {

+      svc->ref_frame_index[cpi->lst_fb_idx] = svc->current_superframe;

+      svc->ref_frame_index[cpi->gld_fb_idx] = svc->current_superframe;

+      svc->ref_frame_index[cpi->alt_fb_idx] = svc->current_superframe;

+    } else {

+      if (cpi->refresh_last_frame)

+        svc->ref_frame_index[cpi->lst_fb_idx] = svc->current_superframe;

+      if (cpi->refresh_golden_frame)

+        svc->ref_frame_index[cpi->gld_fb_idx] = svc->current_superframe;

+      if (cpi->refresh_alt_ref_frame)

+        svc->ref_frame_index[cpi->alt_fb_idx] = svc->current_superframe;

+    }

+  }

 }

 static void loopfilter_frame(VP9_COMP *cpi, VP9_COMMON *cm) {

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

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

@@ -1094,6 +1094,7 @@

                          BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {

   VP9_COMMON *const cm = &cpi->common;

   SPEED_FEATURES *const sf = &cpi->sf;

+  const SVC *const svc = &cpi->svc;

   TileInfo *const tile_info = &tile_data->tile_info;

   MACROBLOCKD *const xd = &x->e_mbd;

   MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;

@@ -1143,6 +1144,7 @@

   int best_pred_sad = INT_MAX;

   int best_early_term = 0;

   int ref_frame_cost[MAX_REF_FRAMES];

+  int svc_force_zero_mode[3] = {0};

 #if CONFIG_VP9_TEMPORAL_DENOISING

   int64_t zero_last_cost_orig = INT64_MAX;

 #endif

@@ -1196,6 +1198,17 @@

   } else {

     usable_ref_frame = GOLDEN_FRAME;

   }

+

+  // If the reference is temporally aligned with current superframe

+  // (e.g., spatial reference within superframe), constrain the inter mode:

+  // for now only test zero motion.

+  if (cpi->use_svc && svc ->force_zero_mode_spatial_ref) {

+    if (svc->ref_frame_index[cpi->lst_fb_idx] == svc->current_superframe)

+      svc_force_zero_mode[LAST_FRAME - 1] = 1;

+    if (svc->ref_frame_index[cpi->gld_fb_idx] == svc->current_superframe)

+      svc_force_zero_mode[GOLDEN_FRAME - 1] = 1;

+  }

+

   for (ref_frame = LAST_FRAME; ref_frame <= usable_ref_frame; ++ref_frame) {

     const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);

@@ -1248,8 +1261,13 @@

       continue;

     ref_frame = ref_mode_set[idx].ref_frame;

-    if (cpi->use_svc)

+    if (cpi->use_svc) {

       ref_frame = ref_mode_set_svc[idx].ref_frame;

+      if (svc_force_zero_mode[ref_frame - 1] &&

+          frame_mv[this_mode][ref_frame].as_int != 0)

+        continue;

+    }

+

     if (!(cpi->ref_frame_flags & flag_list[ref_frame]))

       continue;

     if (const_motion[ref_frame] && this_mode == NEARMV)

--- a/vp9/encoder/vp9_svc_layercontext.c

+++ b/vp9/encoder/vp9_svc_layercontext.c

@@ -25,7 +25,7 @@

   const VP9EncoderConfig *const oxcf = &cpi->oxcf;

   int mi_rows = cpi->common.mi_rows;

   int mi_cols = cpi->common.mi_cols;

-  int sl, tl;

+  int sl, tl, i;

   int alt_ref_idx = svc->number_spatial_layers;

   svc->spatial_layer_id = 0;

@@ -32,6 +32,10 @@

   svc->temporal_layer_id = 0;

   svc->first_spatial_layer_to_encode = 0;

   svc->rc_drop_superframe = 0;

+  svc->force_zero_mode_spatial_ref = 0;

+  svc->current_superframe = 0;

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

+    svc->ref_frame_index[i] = -1;

   if (cpi->oxcf.error_resilient_mode == 0 && cpi->oxcf.pass == 2) {

     if (vpx_realloc_frame_buffer(&cpi->svc.empty_frame.img,

@@ -353,6 +357,8 @@

                               cpi->svc.number_temporal_layers];

   ++lc->current_video_frame_in_layer;

   ++lc->frames_from_key_frame;

+  if (cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1)

+    ++cpi->svc.current_superframe;

 }

 int vp9_is_upper_layer_key_frame(const VP9_COMP *const cpi) {

@@ -542,6 +548,7 @@

 int vp9_one_pass_cbr_svc_start_layer(VP9_COMP *const cpi) {

   int width = 0, height = 0;

   LAYER_CONTEXT *lc = NULL;

+  cpi->svc.force_zero_mode_spatial_ref = 1;

   if (cpi->svc.temporal_layering_mode == VP9E_TEMPORAL_LAYERING_MODE_0212) {

     set_flags_and_fb_idx_for_temporal_mode3(cpi);

--- a/vp9/encoder/vp9_svc_layercontext.h

+++ b/vp9/encoder/vp9_svc_layercontext.h

@@ -83,6 +83,9 @@

   int ext_lst_fb_idx[VPX_MAX_LAYERS];

   int ext_gld_fb_idx[VPX_MAX_LAYERS];

   int ext_alt_fb_idx[VPX_MAX_LAYERS];

+  int ref_frame_index[REF_FRAMES];

+  int force_zero_mode_spatial_ref;

+  int current_superframe;

 } SVC;

 struct VP9_COMP;