]> git.sesse.net Git - x264/blobdiff - encoder/ratecontrol.c
eliminate some divisions
[x264] / encoder / ratecontrol.c
index 5bc3f6181a41480fc0bf1cdc9afd7711c4060149..7054c35ad5dc10b9788e543b79a9f5cf90534e7d 100644 (file)
@@ -31,6 +31,7 @@
 #include "common/common.h"
 #include "common/cpu.h"
 #include "ratecontrol.h"
+#include "me.h"
 
 typedef struct
 {
@@ -49,6 +50,8 @@ typedef struct
     int s_count;
     float blurred_complexity;
     char direct_mode;
+    int refcount[16];
+    int refs;
 } ratecontrol_entry_t;
 
 typedef struct
@@ -291,7 +294,11 @@ int x264_macroblock_tree_read( x264_t *h, x264_frame_t *frame )
             goto fail;
 
         for( i = 0; i < h->mb.i_mb_count; i++ )
+        {
             frame->f_qp_offset[i] = ((float)(int16_t)endian_fix16( rc->qp_buffer[i] )) * (1/256.0);
+            if( h->frames.b_have_lowres )
+                frame->i_inv_qscale_factor[i] = x264_exp2fix8(frame->f_qp_offset[i]);
+        }
     }
     else
         x264_adaptive_quant_frame( h, frame );
@@ -301,6 +308,33 @@ fail:
     return -1;
 }
 
+int x264_reference_build_list_optimal( x264_t *h )
+{
+    ratecontrol_entry_t *rce = h->rc->rce;
+    x264_frame_t *frames[16];
+    int ref, i;
+
+    if( rce->refs != h->i_ref0 )
+        return -1;
+
+    memcpy( frames, h->fref0, sizeof(frames) );
+
+    /* For now don't reorder ref 0; it seems to lower quality
+       in most cases due to skips. */
+    for( ref = 1; ref < h->i_ref0; ref++ )
+    {
+        int max = -1;
+        int bestref = 1;
+        for( i = 1; i < h->i_ref0; i++ )
+            /* Favor lower POC as a tiebreaker. */
+            COPY2_IF_GT( max, rce->refcount[i], bestref, i );
+        rce->refcount[bestref] = -1;
+        h->fref0[ref] = frames[bestref];
+    }
+
+    return 0;
+}
+
 static char *x264_strcat_filename( char *input, char *suffix )
 {
     char *output = x264_malloc( strlen( input ) + strlen( suffix ) + 1 );
@@ -581,6 +615,7 @@ int x264_ratecontrol_new( x264_t *h )
             int e;
             char *next;
             float qp;
+            int ref;
 
             next= strchr(p, ';');
             if(next)
@@ -603,6 +638,20 @@ int x264_ratecontrol_new( x264_t *h )
                    &rce->mv_bits, &rce->misc_bits, &rce->i_count, &rce->p_count,
                    &rce->s_count, &rce->direct_mode);
 
+            p = strstr( p, "ref:" );
+            if( !p )
+                goto parse_error;
+            p += 4;
+            for( ref = 0; ref < 16; ref++ )
+            {
+                if( sscanf( p, " %d", &rce->refcount[ref] ) != 1 )
+                    break;
+                p = strchr( p+1, ' ' );
+                if( !p )
+                    goto parse_error;
+            }
+            rce->refs = ref;
+
             switch(pict_type)
             {
                 case 'I': rce->kept_as_ref = 1;
@@ -614,6 +663,7 @@ int x264_ratecontrol_new( x264_t *h )
             }
             if(e < 10)
             {
+parse_error:
                 x264_log(h, X264_LOG_ERROR, "statistics are damaged at line %d, parser out=%d\n", i, e);
                 return -1;
             }
@@ -988,7 +1038,7 @@ static double predict_row_size( x264_t *h, int y, int qp )
     double pred_t = 0;
     if( h->sh.i_type == SLICE_TYPE_I || qp >= h->fref0[0]->i_row_qp[y] )
     {
-        if( h->sh.i_type != SLICE_TYPE_I
+        if( h->sh.i_type == SLICE_TYPE_P
             && h->fref0[0]->i_type == h->fdec->i_type
             && h->fref0[0]->i_row_satd[y] > 0
             && (abs(h->fref0[0]->i_row_satd[y] - h->fdec->i_row_satd[y]) < h->fdec->i_row_satd[y]/2))
@@ -1046,87 +1096,75 @@ void x264_ratecontrol_mb( x264_t *h, int bits )
 
     h->fdec->i_row_qp[y] = rc->qpm;
 
-    if( h->sh.i_type == SLICE_TYPE_B )
+    update_predictor( rc->row_pred[0], qp2qscale(rc->qpm), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
+    if( h->sh.i_type == SLICE_TYPE_P && rc->qpm < h->fref0[0]->i_row_qp[y] )
     {
-        /* B-frames shouldn't use lower QP than their reference frames.
-         * This code is a bit overzealous in limiting B-frame quantizers, but it helps avoid
-         * underflows due to the fact that B-frames are not explicitly covered by VBV. */
-        if( y < h->sps->i_mb_height-1 )
-        {
-            int i_estimated;
-            int avg_qp = X264_MIN(h->fref0[0]->i_row_qp[y+1], h->fref1[0]->i_row_qp[y+1])
-                       + rc->pb_offset * ((h->fenc->i_type == X264_TYPE_BREF) ? 0.5 : 1);
-            rc->qpm = X264_MIN(X264_MAX( rc->qp, avg_qp), 51); //avg_qp could go higher than 51 due to pb_offset
-            i_estimated = row_bits_so_far(h, y); //FIXME: compute full estimated size
-            if (i_estimated > h->rc->frame_size_planned)
-                x264_ratecontrol_set_estimated_size(h, i_estimated);
-        }
+        double newq = qp2qscale(rc->qpm);
+        double oldq = qp2qscale(h->fref0[0]->i_row_qp[y]);
+        update_predictor( rc->row_pred[1], (1 - newq / oldq) * newq, h->fdec->i_row_satds[0][0][y], h->fdec->i_row_bits[y] );
     }
-    else
+
+    /* tweak quality based on difference from predicted size */
+    if( y < h->sps->i_mb_height-1 )
     {
-        update_predictor( rc->row_pred[0], qp2qscale(rc->qpm), h->fdec->i_row_satd[y], h->fdec->i_row_bits[y] );
-        if( h->sh.i_type != SLICE_TYPE_I && rc->qpm < h->fref0[0]->i_row_qp[y] )
+        int prev_row_qp = h->fdec->i_row_qp[y];
+        int i_qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, h->param.rc.i_qp_max );
+        int i_qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
+
+        /* B-frames shouldn't use lower QP than their reference frames. */
+        if( h->sh.i_type == SLICE_TYPE_B )
         {
-            double newq = qp2qscale(rc->qpm);
-            double oldq = qp2qscale(h->fref0[0]->i_row_qp[y]);
-            update_predictor( rc->row_pred[1], (1 - newq / oldq) * newq, h->fdec->i_row_satds[0][0][y], h->fdec->i_row_bits[y] );
+            i_qp_min = X264_MAX( i_qp_min, X264_MAX( h->fref0[0]->i_row_qp[y+1], h->fref1[0]->i_row_qp[y+1] ) );
+            rc->qpm = X264_MAX( rc->qpm, i_qp_min );
         }
 
-        /* tweak quality based on difference from predicted size */
-        if( y < h->sps->i_mb_height-1 )// && h->stat.i_frame_count[h->sh.i_type] > 0 )
-        {
-            int prev_row_qp = h->fdec->i_row_qp[y];
-            int b0 = predict_row_size_sum( h, y, rc->qpm );
-            int b1 = b0;
-            int i_qp_max = X264_MIN( prev_row_qp + h->param.rc.i_qp_step, h->param.rc.i_qp_max );
-            int i_qp_min = X264_MAX( prev_row_qp - h->param.rc.i_qp_step, h->param.rc.i_qp_min );
-            float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
-
-            /* More threads means we have to be more cautious in letting ratecontrol use up extra bits.
-             * In 2-pass mode we can be more trusting of the planned frame sizes, since they were decided
-             * by actual encoding instead of SATD prediction. */
-            float rc_tol = h->param.rc.b_stat_read ? (buffer_left_planned / rc->buffer_size) * rc->frame_size_planned
-                                                   : (buffer_left_planned / h->param.i_threads);
-            /* Don't modify the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
-            /* area at the top of the frame was measured inaccurately. */
-            if( row_bits_so_far(h,y) < 0.05 * rc->frame_size_planned )
-                return;
-
-            if(h->sh.i_type != SLICE_TYPE_I)
-                rc_tol /= 2;
-
-            if( !rc->b_vbv_min_rate )
-                i_qp_min = X264_MAX( i_qp_min, h->sh.i_qp );
-
-            while( rc->qpm < i_qp_max
-                   && ((b1 > rc->frame_size_planned + rc_tol) ||
-                       (rc->buffer_fill - b1 < buffer_left_planned * 0.5) ||
-                       (b1 > rc->frame_size_planned && rc->qpm < rc->qp_novbv)) )
-            {
-                rc->qpm ++;
-                b1 = predict_row_size_sum( h, y, rc->qpm );
-            }
+        int b0 = predict_row_size_sum( h, y, rc->qpm );
+        int b1 = b0;
+        float buffer_left_planned = rc->buffer_fill - rc->frame_size_planned;
 
-            while( rc->qpm > i_qp_min
-                   && (rc->qpm > h->fdec->i_row_qp[0] || rc->single_frame_vbv)
-                   && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
-                   || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
-            {
-                rc->qpm --;
-                b1 = predict_row_size_sum( h, y, rc->qpm );
-            }
+        /* More threads means we have to be more cautious in letting ratecontrol use up extra bits. */
+        float rc_tol = buffer_left_planned / h->param.i_threads * rc->rate_tolerance;
 
-            /* avoid VBV underflow */
-            while( (rc->qpm < h->param.rc.i_qp_max)
-                   && (rc->buffer_fill - b1 < rc->buffer_rate * 0.05 ) )
-            {
-                rc->qpm ++;
-                b1 = predict_row_size_sum( h, y, rc->qpm );
-            }
+        /* Don't modify the row QPs until a sufficent amount of the bits of the frame have been processed, in case a flat */
+        /* area at the top of the frame was measured inaccurately. */
+        if( row_bits_so_far(h,y) < 0.05 * rc->frame_size_planned )
+            return;
+
+        if( h->sh.i_type != SLICE_TYPE_I )
+            rc_tol /= 2;
+
+        if( !rc->b_vbv_min_rate )
+            i_qp_min = X264_MAX( i_qp_min, h->sh.i_qp );
+
+        while( rc->qpm < i_qp_max
+               && ((b1 > rc->frame_size_planned + rc_tol) ||
+                   (rc->buffer_fill - b1 < buffer_left_planned * 0.5) ||
+                   (b1 > rc->frame_size_planned && rc->qpm < rc->qp_novbv)) )
+        {
+            rc->qpm ++;
+            b1 = predict_row_size_sum( h, y, rc->qpm );
+        }
+
+        while( rc->qpm > i_qp_min
+               && (rc->qpm > h->fdec->i_row_qp[0] || rc->single_frame_vbv)
+               && ((b1 < rc->frame_size_planned * 0.8 && rc->qpm <= prev_row_qp)
+               || b1 < (rc->buffer_fill - rc->buffer_size + rc->buffer_rate) * 1.1) )
+        {
+            rc->qpm --;
+            b1 = predict_row_size_sum( h, y, rc->qpm );
+        }
 
-            x264_ratecontrol_set_estimated_size(h, b1);
+        /* avoid VBV underflow */
+        while( (rc->qpm < h->param.rc.i_qp_max)
+               && (rc->buffer_fill - b1 < rc->buffer_rate * 0.05 ) )
+        {
+            rc->qpm ++;
+            b1 = predict_row_size_sum( h, y, rc->qpm );
         }
+
+        x264_ratecontrol_set_estimated_size(h, b1);
     }
+
     /* loses the fractional part of the frame-wise qp */
     rc->f_qpm = rc->qpm;
 }
@@ -1222,7 +1260,7 @@ int x264_ratecontrol_end( x264_t *h, int bits )
                           dir_avg>0 ? 's' : dir_avg<0 ? 't' : '-' )
                         : '-';
         if( fprintf( rc->p_stat_file_out,
-                 "in:%d out:%d type:%c q:%.2f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c;\n",
+                 "in:%d out:%d type:%c q:%.2f tex:%d mv:%d misc:%d imb:%d pmb:%d smb:%d d:%c ref:",
                  h->fenc->i_frame, h->i_frame,
                  c_type, rc->qpa_rc,
                  h->stat.frame.i_tex_bits,
@@ -1232,7 +1270,19 @@ int x264_ratecontrol_end( x264_t *h, int bits )
                  h->stat.frame.i_mb_count_p,
                  h->stat.frame.i_mb_count_skip,
                  c_direct) < 0 )
-             goto fail;
+            goto fail;
+
+        for( i = 0; i < h->i_ref0; i++ )
+        {
+            int refcount = h->param.b_interlaced ? h->stat.frame.i_mb_count_ref[0][i*2]
+                                                 + h->stat.frame.i_mb_count_ref[0][i*2+1] :
+                                                   h->stat.frame.i_mb_count_ref[0][i];
+            if( fprintf( rc->p_stat_file_out, "%d ", refcount ) < 0 )
+                goto fail;
+        }
+
+        if( fprintf( rc->p_stat_file_out, ";\n" ) < 0 )
+            goto fail;
 
         /* Don't re-write the data in multi-pass mode. */
         if( h->param.rc.b_mb_tree && h->fenc->b_kept_as_ref && !h->param.rc.b_stat_read )
@@ -1646,9 +1696,15 @@ static float rate_estimate_qscale( x264_t *h )
         else
             q += rcc->pb_offset;
 
-        rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref1[h->i_ref1-1]->i_satd );
+        if( rcc->b_2pass && rcc->b_vbv )
+            rcc->frame_size_planned = qscale2bits( &rce, q );
+        else
+            rcc->frame_size_planned = predict_size( rcc->pred_b_from_p, q, h->fref1[h->i_ref1-1]->i_satd );
         x264_ratecontrol_set_estimated_size(h, rcc->frame_size_planned);
-        rcc->last_satd = 0;
+
+        /* For row SATDs */
+        if( rcc->b_vbv )
+            rcc->last_satd = x264_rc_analyse_slice( h );
         return qp2qscale(q);
     }
     else
@@ -1717,7 +1773,7 @@ static float rate_estimate_qscale( x264_t *h )
                     expected_size = qscale2bits(&rce, q);
                     expected_vbv = rcc->buffer_fill + rcc->buffer_rate - expected_size;
                 }
-                rcc->last_satd = x264_stack_align( x264_rc_analyse_slice, h );
+                rcc->last_satd = x264_rc_analyse_slice( h );
             }
             q = x264_clip3f( q, lmin, lmax );
         }
@@ -1735,7 +1791,7 @@ static float rate_estimate_qscale( x264_t *h )
 
             double wanted_bits, overflow=1, lmin, lmax;
 
-            rcc->last_satd = x264_stack_align( x264_rc_analyse_slice, h );
+            rcc->last_satd = x264_rc_analyse_slice( h );
             rcc->short_term_cplxsum *= 0.5;
             rcc->short_term_cplxcount *= 0.5;
             rcc->short_term_cplxsum += rcc->last_satd;