ref: 51319268d6ccf4aca8ec0afefe2a3a41acb31497
parent: c8ea3edb24db14c9a4386d458b0823c297cae5fc
author: Gregory Maxwell <greg@xiph.org>
date: Tue Aug 23 08:51:24 EDT 2011
Add noise shaping dither to opusdec.
--- a/src/opusdec.c
+++ b/src/opusdec.c
@@ -42,6 +42,7 @@
#include "getopt_win.h"
#endif*/
#include <stdlib.h>
+#include <limits.h>
#include <string.h>
#include <opus.h>
@@ -82,6 +83,10 @@
#include "opus_header.h"
#include "speex_resampler.h"
+#define MINI(_a,_b) ((_a)<(_b)?(_a):(_b))
+#define MAXI(_a,_b) ((_a)>(_b)?(_a):(_b))
+#define CLAMPI(_a,_b,_c) (MAXI(_a,MINI(_b,_c)))
+
#define MAX_FRAME_SIZE (2*960*3)
#define readint(buf, base) (((buf[base+3]<<24)&0xff000000)| \
@@ -89,6 +94,93 @@
((buf[base+1]<<8)&0xff00)| \
(buf[base]&0xff))
+typedef struct shapestate shapestate;
+struct shapestate {+ float * b_buf;
+ float * a_buf;
+ int fs;
+ int mute;
+};
+
+static unsigned int rngseed = 22222;
+static inline unsigned int fast_rand() {+ rngseed = (rngseed * 96314165) + 907633515;
+ return rngseed;
+}
+
+/* This implements a 16 bit quantization with full triangular dither
+ and IIR noise shaping. The noise shaping filters were designed by
+ Sebastian Gesemann based on the LAME ATH curves with flattening
+ to limit their peak gain to 20dB.
+ (Everyone elses' noise shaping filters are mildly crazy)
+ The 48kHz version of this filter is just a warped version of the
+ 44.1kHz filter and probably could be improved by shifting the
+ HF shelf up in frequency a little bit since 48k has a bit more
+ room and being more conservative against bat-ears is probably
+ more important than more noise suppression.
+ This process can increase the peak level of the signal (in theory
+ by the peak error of 1.5 +20dB though this much is unobservable rare)
+ so to avoid clipping the signal is attenuated by a couple thousandths
+ of a dB. Initially the approach taken here was to only attenuate by
+ the 99.9th percentile, making clipping rare but not impossible (like
+ SoX) but the limited gain of the filter means that the worst case was
+ only two thousandths of a dB more, so this just uses the worst case.
+ The attenuation is probably also helpful to prevent clipping in the DAC
+ reconstruction filters or downstream resampling in any case.*/
+static inline void shape_dither_toshort(shapestate *_ss, short *_o, float *_i, int _n, int _CC)
+{+ const float gains[3]={32768.f-15.f,32768.f-15.f,32768.f-3.f};+ const float fcoef[3][8] =
+ {+ {2.2374f, -.7339f, -.1251f, -.6033f, 0.9030f, .0116f, -.5853f, -.2571f}, /* 48.0kHz noise shaping filter sd=2.34*/+ {2.2061f, -.4706f, -.2534f, -.6214f, 1.0587f, .0676f, -.6054f, -.2738f}, /* 44.1kHz noise shaping filter sd=2.51*/+ {1.0000f, 0.0000f, 0.0000f, 0.0000f, 0.0000f,0.0000f, 0.0000f, 0.0000f}, /* lowpass noise shaping filter sd=0.65*/+ };
+ int i;
+ int rate=_ss->fs==44100?1:(_ss->fs==48000?0:2);
+ float gain=gains[rate];
+ float *b_buf;
+ float *a_buf;
+ int mute=_ss->mute;
+ b_buf=_ss->b_buf;
+ a_buf=_ss->a_buf;
+ if(mute>64)
+ memset(a_buf,0,sizeof(float)*_CC*4);
+ for(i=0;i<_n;i++)
+ {+ int c;
+ int pos = i*_CC;
+ int silent=1;
+ for(c=0;c<_CC;c++)
+ {+ int j, si;
+ float r,s,err=0;
+ silent&=_i[pos+c]==0;
+ s=_i[pos+c]*gain;
+ for(j=0;j<4;j++)
+ err += fcoef[rate][j]*b_buf[c*4+j] - fcoef[rate][j+4]*a_buf[c*4+j];
+ memmove(&a_buf[c*4+1],&a_buf[c*4],sizeof(float)*3);
+ memmove(&b_buf[c*4+1],&b_buf[c*4],sizeof(float)*3);
+ a_buf[c*4]=err;
+ s = s - err;
+ r=(float)fast_rand()*(1/(float)UINT_MAX) - (float)fast_rand()*(1/(float)UINT_MAX);
+ if (mute>16)r=0;
+ /*Clamp in float out of paranoia that the input will be >96dBFS and wrap if the
+ integer is clamped.*/
+ _o[pos+c] = si = lrintf(fmaxf(-32768,fminf(s + r,32767)));
+ /*Including clipping in the noise shaping is generally disastrous:
+ the futile effort to restore the clipped energy results in more clipping.
+ However, small amounts-- at the level which could normally be created by
+ dither and rounding-- are harmless and can even reduce clipping somewhat
+ due to the clipping sometimes reducing the dither+rounding error.*/
+ b_buf[c*4] = (mute>16)?0:fmaxf(-1.5f,fminf(si - s,1.5f));
+ }
+ mute++;
+ if(!silent)mute=0;
+ }
+ _ss->mute=MINI(mute,960);
+}
+
static void print_comments(char *comments, int length)
{char *c=comments;
@@ -273,6 +365,7 @@
printf (" --mono Force decoding in mono\n"); printf (" --stereo Force decoding in stereo\n"); printf (" --rate n Force decoding at sampling rate n Hz\n");+ printf (" --no-dither Do not dither 16-bit output\n"); printf (" --packet-loss n Simulate n %% random packet loss\n"); printf (" -V Verbose mode (show bit-rate)\n"); printf (" -h, --help This help\n");@@ -282,13 +375,13 @@
void version(void)
{- printf ("opusenc (based on %s)\n",opus_get_version_string());+ printf ("opusdec (based on %s)\n",opus_get_version_string()); printf ("Copyright (C) 2008-2011 Jean-Marc Valin\n");}
void version_short(void)
{- printf ("opusenc (based on %s)\n",opus_get_version_string());+ printf ("opusdec (based on %s)\n",opus_get_version_string()); printf ("Copyright (C) 2008-2011 Jean-Marc Valin\n");}
@@ -327,7 +420,7 @@
printf("Playback gain: %f (%f dB)\n", *gain, header.gain/256.);if (!quiet)
{- fprintf (stderr, "Decoding %d Hz audio in", *rate);
+ fprintf (stderr, "Decoding %d Hz audio", *rate);
if (*channels==1)
fprintf (stderr, " (mono");
@@ -339,7 +432,7 @@
return st;
}
-void audio_write(float *pcm, int channels, int frame_size, FILE *fout, SpeexResamplerState *resampler, int *skip, int file)
+void audio_write(float *pcm, int channels, int frame_size, FILE *fout, SpeexResamplerState *resampler, int *skip, shapestate *shapemem, int file)
{int i,tmp_skip;
unsigned out_len;
@@ -370,14 +463,17 @@
}
/*Convert to short and save to output file*/
- /*FIXME: This should dither for integer output*/
- if (file){+ if (shapemem){+ shape_dither_toshort(shapemem,out,output,out_len,channels);
+ }else{for (i=0;i<out_len*channels;i++)
- out[i]=le_short((short)lrintf(fmax(fmin(output[i]*32768.f+0.5f,32767),-32768)));
- } else {+ out[i]=(short)lrintf(fmax(-32768,fmin(output[i]*32768.f,32767)));
+ }
+ if ((le_short(1)!=1)&&file){for (i=0;i<out_len*channels;i++)
- out[i]=(short)lrintf(fmax(fmin(output[i]*32768.f+0.5f,32767),-32768));
+ out[i]=le_short(out[i]);
}
+
fwrite(out+tmp_skip*channels, 2, (out_len-tmp_skip)*channels, fout);
} while (frame_size != 0);
}
@@ -405,6 +501,7 @@
{"rate", required_argument, NULL, 0}, {"mono", no_argument, NULL, 0}, {"stereo", no_argument, NULL, 0},+ {"no-dither", no_argument, NULL, 0}, {"packet-loss", required_argument, NULL, 0}, {0, 0, 0, 0}};
@@ -423,8 +520,15 @@
int wav_format=0;
int preskip=0;
int opus_serialno = -1;
+ int dither=1;
+ shapestate shapemem;
SpeexResamplerState *resampler=NULL;
float gain=1;
+
+ shapemem.a_buf=0;
+ shapemem.b_buf=0;
+ shapemem.mute=960;
+ shapemem.fs=0;
enh_enabled = 1;
@@ -460,6 +564,9 @@
} else if (strcmp(long_options[option_index].name,"stereo")==0)
{channels=2;
+ } else if (strcmp(long_options[option_index].name,"no-dither")==0)
+ {+ dither=0;
} else if (strcmp(long_options[option_index].name,"rate")==0)
{rate=atoi (optarg);
@@ -560,6 +667,13 @@
if (packet_count==0)
{st = process_header(&op, &rate, &channels, &preskip, &gain, quiet);
+ if(shapemem.a_buf)
+ free(shapemem.a_buf);
+ if(shapemem.b_buf)
+ free(shapemem.b_buf);
+ shapemem.a_buf=calloc(channels,sizeof(float)*4);
+ shapemem.b_buf=calloc(channels,sizeof(float)*4);
+ shapemem.fs=rate;
/* Converting preskip to output sampling rate */
preskip = preskip*(rate/48000.);
if (!st)
@@ -626,7 +740,7 @@
if (truncate > frame_size)
truncate = frame_size;
new_frame_size = frame_size - truncate;
- audio_write(output, channels, new_frame_size, fout, resampler, &preskip, strlen(outFile)==0);
+ audio_write(output, channels, new_frame_size, fout, resampler, &preskip, dither?&shapemem:0, strlen(outFile)==0);
audio_size+=sizeof(short)*new_frame_size*channels;
}
}
@@ -653,7 +767,7 @@
int tmp = drain;
if (tmp > 100)
tmp = 100;
- audio_write(zeros, channels, tmp, fout, resampler, NULL, strlen(outFile)==0);
+ audio_write(zeros, channels, tmp, fout, resampler, NULL, &shapemem, strlen(outFile)==0);
drain -= tmp;
} while (drain>0);
}