Aggressive VBR method.

VBR method that allows a wider Q range for the first normal frame
in each ARF group and then centers the min - max range for the rest of
the arf group on the chosen Q value for that first frame.

This allows for quite rapid adjustment of the active Q range even if the
initial estimate is poor.

In some cases where the ARF frames themselves are tending to
undershoot but the normal frames are overshooting this can still give
net undershoot. This can be corrected by allowing a larger Q delta for
arf frames but is usually is a sign that the allocation to the arfs was to
high.

Change-Id: Icec87758925d8f7aeb2dca29aac0ff9496237469

diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c
index 0dce44c..293d7e0 100644 (file)
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -3317,6 +3317,9 @@ static void encode_with_recode_loop(VP9_COMP *cpi, size_t *size,
   int frame_under_shoot_limit;
   int q = 0, q_low = 0, q_high = 0;
   int enable_acl;
+#ifdef AGRESSIVE_VBR
+  int qrange_adj = 1;
+#endif
 
   set_size_independent_vars(cpi);
 
@@ -3332,6 +3335,17 @@ static void encode_with_recode_loop(VP9_COMP *cpi, size_t *size,
     if (loop_count == 0 || cpi->resize_pending != 0) {
       set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
 
+#ifdef AGRESSIVE_VBR
+      if (two_pass_first_group_inter(cpi)) {
+        // Adjustment limits for min and max q
+        qrange_adj = VPXMAX(1, (top_index - bottom_index) / 2);
+
+        bottom_index = VPXMAX(bottom_index - qrange_adj / 2,
+                              cpi->oxcf.best_allowed_q);
+        top_index = VPXMIN(cpi->oxcf.worst_allowed_q,
+                           top_index + qrange_adj / 2);
+      }
+#endif
       // TODO(agrange) Scale cpi->max_mv_magnitude if frame-size has changed.
       set_mv_search_params(cpi);
 
@@ -3592,6 +3606,13 @@ static void encode_with_recode_loop(VP9_COMP *cpi, size_t *size,
       if (loop || !enable_acl) restore_coding_context(cpi);
   } while (loop);
 
+#ifdef AGRESSIVE_VBR
+  if (two_pass_first_group_inter(cpi)) {
+    cpi->twopass.active_worst_quality =
+       VPXMIN(q + qrange_adj, cpi->oxcf.worst_allowed_q);
+  }
+#endif
+
   if (enable_acl) {
     vp9_encode_frame(cpi);
     vpx_clear_system_state();

diff --git a/vp9/encoder/vp9_firstpass.c b/vp9/encoder/vp9_firstpass.c
index 59dd536..8250ff3 100644 (file)
--- a/vp9/encoder/vp9_firstpass.c
+++ b/vp9/encoder/vp9_firstpass.c
@@ -2463,7 +2463,11 @@ static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
   }
 
   // Limit maximum boost based on interval length.
+#ifdef AGRESSIVE_VBR
+  rc->gfu_boost = VPXMIN((int)rc->gfu_boost, i * 140);
+#else
   rc->gfu_boost = VPXMIN((int)rc->gfu_boost, i * 200);
+#endif
 
   // Set the interval until the next gf.
   rc->baseline_gf_interval = i - (is_key_frame || rc->source_alt_ref_pending);
@@ -2666,11 +2670,17 @@ static int test_candidate_kf(TWO_PASS *twopass,
 }
 
 #define FRAMES_TO_CHECK_DECAY 8
-#define KF_MAX_FRAME_BOOST 96.0
 #define MIN_KF_TOT_BOOST 300
-#define MAX_KF_TOT_BOOST 5400
 #define KF_BOOST_SCAN_MAX_FRAMES 32
 
+#ifdef AGRESSIVE_VBR
+#define KF_MAX_FRAME_BOOST 80.0
+#define MAX_KF_TOT_BOOST 4800
+#else
+#define KF_MAX_FRAME_BOOST 96.0
+#define MAX_KF_TOT_BOOST 5400
+#endif
+
 static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
   int i, j;
   RATE_CONTROL *const rc = &cpi->rc;

diff --git a/vp9/encoder/vp9_ratectrl.c b/vp9/encoder/vp9_ratectrl.c
index f781622..ba22938 100644 (file)
--- a/vp9/encoder/vp9_ratectrl.c
+++ b/vp9/encoder/vp9_ratectrl.c
@@ -93,10 +93,17 @@ static int inter_minq_12[QINDEX_RANGE];
 static int rtc_minq_12[QINDEX_RANGE];
 #endif
 
+#ifdef AGRESSIVE_VBR
+static int gf_high = 2400;
+static int gf_low = 400;
+static int kf_high = 4000;
+static int kf_low = 400;
+#else
 static int gf_high = 2000;
 static int gf_low = 400;
 static int kf_high = 5000;
 static int kf_low = 400;
+#endif
 
 // Functions to compute the active minq lookup table entries based on a
 // formulaic approach to facilitate easier adjustment of the Q tables.
@@ -126,9 +133,14 @@ static void init_minq_luts(int *kf_low_m, int *kf_high_m, int *arfgf_low,
     const double maxq = vp9_convert_qindex_to_q(i, bit_depth);
     kf_low_m[i] = get_minq_index(maxq, 0.000001, -0.0004, 0.150, bit_depth);
     kf_high_m[i] = get_minq_index(maxq, 0.0000021, -0.00125, 0.55, bit_depth);
+#ifdef AGRESSIVE_VBR
+    arfgf_low[i] = get_minq_index(maxq, 0.0000015, -0.0009, 0.275, bit_depth);
+    inter[i] = get_minq_index(maxq, 0.00000271, -0.00113, 0.80, bit_depth);
+#else
     arfgf_low[i] = get_minq_index(maxq, 0.0000015, -0.0009, 0.30, bit_depth);
-    arfgf_high[i] = get_minq_index(maxq, 0.0000021, -0.00125, 0.55, bit_depth);
     inter[i] = get_minq_index(maxq, 0.00000271, -0.00113, 0.70, bit_depth);
+#endif
+    arfgf_high[i] = get_minq_index(maxq, 0.0000021, -0.00125, 0.55, bit_depth);
     rtc[i] = get_minq_index(maxq, 0.00000271, -0.00113, 0.70, bit_depth);
   }
 }

diff --git a/vp9/encoder/vp9_ratectrl.h b/vp9/encoder/vp9_ratectrl.h
index 70aef03..a3f73fb 100644 (file)
--- a/vp9/encoder/vp9_ratectrl.h
+++ b/vp9/encoder/vp9_ratectrl.h
@@ -21,6 +21,9 @@
 extern "C" {
 #endif
 
+// Used to control agressive VBR mode.
+// #define AGRESSIVE_VBR 1
+
 // Bits Per MB at different Q (Multiplied by 512)
 #define BPER_MB_NORMBITS 9