shithub: libvpx

Download patch

ref: e51c9e39bcfcf26b9031845ff7767050bcb3b059
parent: 5222a928b04d80fdc88d0d566aa0cf931c426bfb
parent: b6f7d694647a587dacbab1dc65d40c4630aa9632
author: Yaowu Xu <yaowu@google.com>
date: Mon May 14 16:56:34 EDT 2018

Merge "Fixes for consistent encoding across recodes of a frame"

--- a/vp9/encoder/vp9_encodeframe.c
+++ b/vp9/encoder/vp9_encodeframe.c
@@ -4663,6 +4663,9 @@
         for (i = 0; i < BLOCK_SIZES; ++i) {
           for (j = 0; j < MAX_MODES; ++j) {
             tile_data->thresh_freq_fact[i][j] = RD_THRESH_INIT_FACT;
+#if CONSISTENT_RECODE_STATE
+            tile_data->thresh_freq_fact_prev[i][j] = RD_THRESH_INIT_FACT;
+#endif
             tile_data->mode_map[i][j] = j;
           }
         }
@@ -4787,7 +4790,9 @@
   x->fwd_txfm4x4 = xd->lossless ? vp9_fwht4x4 : vpx_fdct4x4;
 #endif  // CONFIG_VP9_HIGHBITDEPTH
   x->inv_txfm_add = xd->lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
-
+#if CONSISTENT_RECODE_STATE
+  x->optimize = sf->optimize_coefficients == 1 && cpi->oxcf.pass != 1;
+#endif
   if (xd->lossless) x->optimize = 0;
 
   cm->tx_mode = select_tx_mode(cpi, xd);
@@ -4912,8 +4917,47 @@
   return sum_delta / (cm->mi_rows * cm->mi_cols);
 }
 
+#if CONSISTENT_RECODE_STATE
+static void restore_encode_params(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  const int tile_cols = 1 << cm->log2_tile_cols;
+  const int tile_rows = 1 << cm->log2_tile_rows;
+  int tile_col, tile_row;
+  int i, j;
+  RD_OPT *rd_opt = &cpi->rd;
+  for (i = 0; i < MAX_REF_FRAMES; i++) {
+    for (j = 0; j < REFERENCE_MODES; j++)
+      rd_opt->prediction_type_threshes[i][j] =
+          rd_opt->prediction_type_threshes_prev[i][j];
+
+    for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; j++)
+      rd_opt->filter_threshes[i][j] = rd_opt->filter_threshes_prev[i][j];
+  }
+
+  if (cpi->tile_data != NULL) {
+    for (tile_row = 0; tile_row < tile_rows; ++tile_row)
+      for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
+        TileDataEnc *tile_data =
+            &cpi->tile_data[tile_row * tile_cols + tile_col];
+        for (i = 0; i < BLOCK_SIZES; ++i) {
+          for (j = 0; j < MAX_MODES; ++j) {
+            tile_data->thresh_freq_fact[i][j] =
+                tile_data->thresh_freq_fact_prev[i][j];
+          }
+        }
+      }
+  }
+
+  cm->interp_filter = cpi->sf.default_interp_filter;
+}
+#endif
+
 void vp9_encode_frame(VP9_COMP *cpi) {
   VP9_COMMON *const cm = &cpi->common;
+
+#if CONSISTENT_RECODE_STATE
+  restore_encode_params(cpi);
+#endif
 
   // In the longer term the encoder should be generalized to match the
   // decoder such that we allow compound where one of the 3 buffers has a
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -3579,6 +3579,39 @@
   set_ref_ptrs(cm, xd, LAST_FRAME, LAST_FRAME);
 }
 
+#if CONSISTENT_RECODE_STATE
+static void save_encode_params(VP9_COMP *cpi) {
+  VP9_COMMON *const cm = &cpi->common;
+  const int tile_cols = 1 << cm->log2_tile_cols;
+  const int tile_rows = 1 << cm->log2_tile_rows;
+  int tile_col, tile_row;
+  int i, j;
+  RD_OPT *rd_opt = &cpi->rd;
+  for (i = 0; i < MAX_REF_FRAMES; i++) {
+    for (j = 0; j < REFERENCE_MODES; j++)
+      rd_opt->prediction_type_threshes_prev[i][j] =
+          rd_opt->prediction_type_threshes[i][j];
+
+    for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; j++)
+      rd_opt->filter_threshes_prev[i][j] = rd_opt->filter_threshes[i][j];
+  }
+
+  if (cpi->tile_data != NULL) {
+    for (tile_row = 0; tile_row < tile_rows; ++tile_row)
+      for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
+        TileDataEnc *tile_data =
+            &cpi->tile_data[tile_row * tile_cols + tile_col];
+        for (i = 0; i < BLOCK_SIZES; ++i) {
+          for (j = 0; j < MAX_MODES; ++j) {
+            tile_data->thresh_freq_fact_prev[i][j] =
+                tile_data->thresh_freq_fact[i][j];
+          }
+        }
+      }
+  }
+}
+#endif
+
 static void encode_without_recode_loop(VP9_COMP *cpi, size_t *size,
                                        uint8_t *dest) {
   VP9_COMMON *const cm = &cpi->common;
@@ -4610,6 +4643,10 @@
 #if CONFIG_INTERNAL_STATS
   memset(cpi->mode_chosen_counts, 0,
          MAX_MODES * sizeof(*cpi->mode_chosen_counts));
+#endif
+#if CONSISTENT_RECODE_STATE
+  // Backup to ensure consistency between recodes
+  save_encode_params(cpi);
 #endif
 
   if (cpi->sf.recode_loop == DISALLOW_RECODE) {
--- a/vp9/encoder/vp9_encoder.h
+++ b/vp9/encoder/vp9_encoder.h
@@ -282,6 +282,9 @@
 typedef struct TileDataEnc {
   TileInfo tile_info;
   int thresh_freq_fact[BLOCK_SIZES][MAX_MODES];
+#if CONSISTENT_RECODE_STATE
+  int thresh_freq_fact_prev[BLOCK_SIZES][MAX_MODES];
+#endif
   int8_t mode_map[BLOCK_SIZES][MAX_MODES];
   FIRSTPASS_DATA fp_data;
   VP9RowMTSync row_mt_sync;
--- a/vp9/encoder/vp9_rd.h
+++ b/vp9/encoder/vp9_rd.h
@@ -23,6 +23,9 @@
 extern "C" {
 #endif
 
+// This macro defines the control for consistent recode behaviour
+#define CONSISTENT_RECODE_STATE 0
+
 #define RDDIV_BITS 7
 #define RD_EPB_SHIFT 6
 
@@ -108,7 +111,11 @@
   int64_t prediction_type_threshes[MAX_REF_FRAMES][REFERENCE_MODES];
 
   int64_t filter_threshes[MAX_REF_FRAMES][SWITCHABLE_FILTER_CONTEXTS];
+#if CONSISTENT_RECODE_STATE
+  int64_t prediction_type_threshes_prev[MAX_REF_FRAMES][REFERENCE_MODES];
 
+  int64_t filter_threshes_prev[MAX_REF_FRAMES][SWITCHABLE_FILTER_CONTEXTS];
+#endif
   int RDMULT;
   int RDDIV;
 } RD_OPT;
--- a/vp9/encoder/vp9_rdopt.c
+++ b/vp9/encoder/vp9_rdopt.c
@@ -3612,9 +3612,13 @@
   }
 
   if (best_mode_index < 0 || best_rd >= best_rd_so_far) {
-    // If adaptive interp filter is enabled, then the current leaf node of 8x8
-    // data is needed for sub8x8. Hence preserve the context.
+// If adaptive interp filter is enabled, then the current leaf node of 8x8
+// data is needed for sub8x8. Hence preserve the context.
+#if CONSISTENT_RECODE_STATE
+    if (bsize == BLOCK_8X8) ctx->mic = *xd->mi[0];
+#else
     if (cpi->row_mt && bsize == BLOCK_8X8) ctx->mic = *xd->mi[0];
+#endif
     rd_cost->rate = INT_MAX;
     rd_cost->rdcost = INT64_MAX;
     return;