Fix high bit depth lookahead cost compensation algorithm

[x264] / encoder / ratecontrol.c

diff --git a/encoder/ratecontrol.c b/encoder/ratecontrol.c
index 363e81ece0f5da99b608ff778462b97c6de86970..8b8e76e643c35370f802620d436307b7b990b590 100644 (file)
--- a/encoder/ratecontrol.c
+++ b/encoder/ratecontrol.c
@@ -1,7 +1,7 @@
 /*****************************************************************************
  * ratecontrol.c: ratecontrol
  *****************************************************************************
- * Copyright (C) 2005-2014 x264 project
+ * Copyright (C) 2005-2015 x264 project
  *
  * Authors: Loren Merritt <lorenm@u.washington.edu>
  *          Michael Niedermayer <michaelni@gmx.at>
@@ -201,11 +201,11 @@ static void update_predictor( predictor_t *p, float q, float var, float bits );
  */
 static inline float qp2qscale( float qp )
 {
-    return 0.85f * powf( 2.0f, ( qp - 12.0f ) / 6.0f );
+    return 0.85f * powf( 2.0f, ( qp - (12.0f + QP_BD_OFFSET) ) / 6.0f );
 }
 static inline float qscale2qp( float qscale )
 {
-    return 12.0f + 6.0f * log2f( qscale/0.85f );
+    return (12.0f + QP_BD_OFFSET) + 6.0f * log2f( qscale/0.85f );
 }
 
 /* Texture bitrate is not quite inversely proportional to qscale,
@@ -302,10 +302,6 @@ static NOINLINE uint32_t x264_ac_energy_mb( x264_t *h, int mb_x, int mb_y, x264_
 
 void x264_adaptive_quant_frame( x264_t *h, x264_frame_t *frame, float *quant_offsets )
 {
-    /* constants chosen to result in approximately the same overall bitrate as without AQ.
-     * FIXME: while they're written in 5 significant digits, they're only tuned to 2. */
-    float strength;
-    float avg_adj = 0.f;
     /* Initialize frame stats */
     for( int i = 0; i < 3; i++ )
     {
@@ -349,23 +345,30 @@ void x264_adaptive_quant_frame( x264_t *h, x264_frame_t *frame, float *quant_off
     /* Actual adaptive quantization */
     else
     {
-        if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
+        /* constants chosen to result in approximately the same overall bitrate as without AQ.
+         * FIXME: while they're written in 5 significant digits, they're only tuned to 2. */
+        float strength;
+        float avg_adj = 0.f;
+        float bias_strength = 0.f;
+
+        if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE || h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE_BIASED )
         {
-            float bit_depth_correction = powf(1 << (BIT_DEPTH-8), 0.5f);
+            float bit_depth_correction = 1.f / (1 << (2*(BIT_DEPTH-8)));
             float avg_adj_pow2 = 0.f;
             for( int mb_y = 0; mb_y < h->mb.i_mb_height; mb_y++ )
                 for( int mb_x = 0; mb_x < h->mb.i_mb_width; mb_x++ )
                 {
                     uint32_t energy = x264_ac_energy_mb( h, mb_x, mb_y, frame );
-                    float qp_adj = powf( energy + 1, 0.125f );
+                    float qp_adj = powf( energy * bit_depth_correction + 1, 0.125f );
                     frame->f_qp_offset[mb_x + mb_y*h->mb.i_mb_stride] = qp_adj;
                     avg_adj += qp_adj;
                     avg_adj_pow2 += qp_adj * qp_adj;
                 }
             avg_adj /= h->mb.i_mb_count;
             avg_adj_pow2 /= h->mb.i_mb_count;
-            strength = h->param.rc.f_aq_strength * avg_adj / bit_depth_correction;
-            avg_adj = avg_adj - 0.5f * (avg_adj_pow2 - (14.f * bit_depth_correction)) / avg_adj;
+            strength = h->param.rc.f_aq_strength * avg_adj;
+            avg_adj = avg_adj - 0.5f * (avg_adj_pow2 - 14.f) / avg_adj;
+            bias_strength = h->param.rc.f_aq_strength;
         }
         else
             strength = h->param.rc.f_aq_strength * 1.0397f;
@@ -375,7 +378,12 @@ void x264_adaptive_quant_frame( x264_t *h, x264_frame_t *frame, float *quant_off
             {
                 float qp_adj;
                 int mb_xy = mb_x + mb_y*h->mb.i_mb_stride;
-                if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
+                if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE_BIASED )
+                {
+                    qp_adj = frame->f_qp_offset[mb_xy];
+                    qp_adj = strength * (qp_adj - avg_adj) + bias_strength * (1.f - 14.f / (qp_adj * qp_adj));
+                }
+                else if( h->param.rc.i_aq_mode == X264_AQ_AUTOVARIANCE )
                 {
                     qp_adj = frame->f_qp_offset[mb_xy];
                     qp_adj = strength * (qp_adj - avg_adj);
@@ -820,10 +828,11 @@ int x264_ratecontrol_new( x264_t *h )
     h->mb.ip_offset = rc->ip_offset + 0.5;
 
     rc->lstep = pow( 2, h->param.rc.i_qp_step / 6.0 );
-    rc->last_qscale = qp2qscale( 26 );
+    rc->last_qscale = qp2qscale( 26 + QP_BD_OFFSET );
     int num_preds = h->param.b_sliced_threads * h->param.i_threads + 1;
     CHECKED_MALLOC( rc->pred, 5 * sizeof(predictor_t) * num_preds );
     CHECKED_MALLOC( rc->pred_b_from_p, sizeof(predictor_t) );
+    static const float pred_coeff_table[3] = { 1.0, 1.0, 1.5 };
     for( int i = 0; i < 3; i++ )
     {
         rc->last_qscale_for[i] = qp2qscale( ABR_INIT_QP );
@@ -831,8 +840,8 @@ int x264_ratecontrol_new( x264_t *h )
         rc->lmax[i] = qp2qscale( h->param.rc.i_qp_max );
         for( int j = 0; j < num_preds; j++ )
         {
-            rc->pred[i+j*5].coeff_min = 2.0 / 4;
-            rc->pred[i+j*5].coeff = 2.0;
+            rc->pred[i+j*5].coeff_min = pred_coeff_table[i] / 2;
+            rc->pred[i+j*5].coeff = pred_coeff_table[i];
             rc->pred[i+j*5].count = 1.0;
             rc->pred[i+j*5].decay = 0.5;
             rc->pred[i+j*5].offset = 0.0;
@@ -846,7 +855,11 @@ int x264_ratecontrol_new( x264_t *h )
             rc->row_preds[i][j].offset = 0.0;
         }
     }
-    *rc->pred_b_from_p = rc->pred[0];
+    rc->pred_b_from_p->coeff_min = 0.5 / 2;
+    rc->pred_b_from_p->coeff = 0.5;
+    rc->pred_b_from_p->count = 1.0;
+    rc->pred_b_from_p->decay = 0.5;
+    rc->pred_b_from_p->offset = 0.0;
 
     if( parse_zones( h ) < 0 )
     {
@@ -1011,7 +1024,7 @@ int x264_ratecontrol_new( x264_t *h )
         {
             ratecontrol_entry_t *rce = &rc->entry[i];
             rce->pict_type = SLICE_TYPE_P;
-            rce->qscale = rce->new_qscale = qp2qscale( 20 );
+            rce->qscale = rce->new_qscale = qp2qscale( 20 + QP_BD_OFFSET );
             rce->misc_bits = rc->nmb + 10;
             rce->new_qp = 0;
         }
@@ -1573,7 +1586,7 @@ int x264_ratecontrol_mb( x264_t *h, int bits )
     h->fdec->f_row_qscale[y] = qscale;
 
     update_predictor( &rc->row_pred[0], qscale, h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
-    if( h->sh.i_type == SLICE_TYPE_P && rc->qpm < h->fref[0][0]->f_row_qp[y] )
+    if( h->sh.i_type != SLICE_TYPE_I && rc->qpm < h->fref[0][0]->f_row_qp[y] )
         update_predictor( &rc->row_pred[1], qscale, h->fdec->i_row_satds[0][0][y], h->fdec->i_row_bits[y] );
 
     /* update ratecontrol per-mbpair in MBAFF */
@@ -2076,7 +2089,8 @@ static void update_predictor( predictor_t *p, float q, float var, float bits )
     if( var < 10 )
         return;
     float old_coeff = p->coeff / p->count;
-    float new_coeff = X264_MAX( bits*q / var, p->coeff_min );
+    float old_offset = p->offset / p->count;
+    float new_coeff = X264_MAX( (bits*q - old_offset) / var, p->coeff_min );
     float new_coeff_clipped = x264_clip3f( new_coeff, old_coeff/range, old_coeff*range );
     float new_offset = bits*q - new_coeff_clipped * var;
     if( new_offset >= 0 )