[xiph-commits] r16192 - branches/theora-thusnelda/lib/enc

[tterribe at svn.xiph.org]

Author: tterribe
Date: 2009-07-03 14:07:46 -0700 (Fri, 03 Jul 2009)
New Revision: 16192

Modified:
   branches/theora-thusnelda/lib/enc/rate.c
Log:
Change the quantizer->lambda mapping.
The new model was built solely considering AC coefficient statistics, and
 ignores DC, which we do not subject to RDO anyway.
In the default VP3 quantization matrices, DC is scaled at a lesser rate than
 AC, and this was playing a significant role in the mapping, especially at low
 rates, where DC-only blocks are much more common.
The new model is quadratic, like the models derived for most other codecs, and
 unlike the previous model, which used an exponent of 1.5.
It produces much higher lambda values at lower rates, and slightly smaller
 values at higher rates, which extends the bitrate range of the codec on both
 ends for a given set of quantization matrices
This improves rate control behavior, especially on the low end, and seems to
 translate into smoother degredation and less blocking.
Overall, it degrades PSNR by slightly more than 0.01 dB for equivalent rates
 (e.g., not much).
The average PSNR on longer sequences with rate control enabled may degrade
 further, but this is due to abililty to hit much lower rates (and thus stay
 within buffer limits more reliably).


Modified: branches/theora-thusnelda/lib/enc/rate.c
===================================================================
--- branches/theora-thusnelda/lib/enc/rate.c	2009-07-02 23:08:59 UTC (rev 16191)
+++ branches/theora-thusnelda/lib/enc/rate.c	2009-07-03 21:07:46 UTC (rev 16192)
@@ -21,28 +21,26 @@
 
 
 void oc_enc_calc_lambda(oc_enc_ctx *_enc,int _frame_type){
-  ogg_int64_t l;
+  ogg_int64_t lq;
   int         qi;
-  qi=_enc->state.qis[0];
   /*For now, lambda is fixed depending on the qi value and frame type:
-      lambda=scale[qti]*(qavg[qti][qi]**1.5),
-     where scale={2.25,1.125}.
-    A more adaptive scheme might perform better, but Theora's behavior does not
-     seem to conform to existing models in the literature.*/
+      lambda=qscale*(qavg[qti][qi]**2),
+     where qscale=0.31.
+    This was derived by exhaustively searching for the optimal quantizer for
+     the AC coefficients in each block from a number of test sequences for a
+     number of fixed lambda values and fitting the peaks of the resulting
+     histograms (on the log(qavg) scale).
+    The same model applies to both inter and intra frames.
+    A more adaptive scheme might perform better.*/
+  qi=_enc->state.qis[0];
   /*If rate control is active, use the lambda for the _target_ quantizer.
     This allows us to scale to rates slightly lower than we'd normally be able
      to reach, and give the rate control a semblance of "fractional qi"
      precision.*/
-  if(_enc->state.info.target_bitrate>0)l=_enc->rc.log_qtarget;
-  else l=_enc->log_qavg[_frame_type][qi];
-  /*Raise to the 1.5 power.*/
-  l+=l>>1;
-  /*Multiply by 1.125.*/
-  l+=0x00570068E7EF5A1ELL;
-  /*And multiply by an extra factor of 2 for INTRA frames.*/
-  if(!_frame_type)l+=OC_Q57(1);
-  /*The upper bound here is 0x48000.*/
-  _enc->lambda=(int)oc_bexp64(l);
+  if(_enc->state.info.target_bitrate>0)lq=_enc->rc.log_qtarget;
+  else lq=_enc->log_qavg[_frame_type][qi];
+  /*The resulting lambda value is less than 0x500000.*/
+  _enc->lambda=(int)oc_bexp64(2*lq-0x3611B1986AB1180LL);
 }
 
 
@@ -61,8 +59,8 @@
   }
   else if(_rc->bits_per_frame<32)_rc->bits_per_frame=32;
   /*The buffer size is set equal to the keyframe interval, clamped to the range
-     [8,256] frames.
-    The 8 frame minimum gives us some chance to distribute bit estimation
+     [12,256] frames.
+    The 12 frame minimum gives us some chance to distribute bit estimation
      errors.
     The 256 frame maximum means we'll require 8-10 seconds of pre-buffering at
      24-30 fps, which is not unreasonable.*/