1 /********************************************************************
2  *                                                                  *
3  * THIS FILE IS PART OF THE OggVorbis SOFTWARE CODEC SOURCE CODE.   *
4  * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
5  * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
6  * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
7  *                                                                  *
8  * THE OggVorbis SOURCE CODE IS (C) COPYRIGHT 1994-2009             *
9  * by the Xiph.Org Foundation http://www.xiph.org/                  *
10  *                                                                  *
11  ********************************************************************
12 
13   function: LSP (also called LSF) conversion routines
14   last mod: $Id: lsp.c 16227 2009-07-08 06:58:46Z xiphmont $
15 
16   The LSP generation code is taken (with minimal modification and a
17   few bugfixes) from "On the Computation of the LSP Frequencies" by
18   Joseph Rothweiler (see http://www.rothweiler.us for contact info).
19   The paper is available at:
20 
21   http://www.myown1.com/joe/lsf
22 
23  ********************************************************************/
24 
25 /* Note that the lpc-lsp conversion finds the roots of polynomial with
26    an iterative root polisher (CACM algorithm 283).  It *is* possible
27    to confuse this algorithm into not converging; that should only
28    happen with absurdly closely spaced roots (very sharp peaks in the
29    LPC f response) which in turn should be impossible in our use of
30    the code.  If this *does* happen anyway, it's a bug in the floor
31    finder; find the cause of the confusion (probably a single bin
32    spike or accidental near-float-limit resolution problems) and
33    correct it. */
34 
35 #include <math.h>
36 #include <string.h>
37 #include <stdlib.h>
38 #include "lsp.h"
39 #include "os.h"
40 #include "misc.h"
41 #include "lookup.h"
42 #include "scales.h"
43 
44 /* three possible LSP to f curve functions; the exact computation
45    (float), a lookup based float implementation, and an integer
46    implementation.  The float lookup is likely the optimal choice on
47    any machine with an FPU.  The integer implementation is *not* fixed
48    point (due to the need for a large dynamic range and thus a
49    seperately tracked exponent) and thus much more complex than the
50    relatively simple float implementations. It's mostly for future
51    work on a fully fixed point implementation for processors like the
52    ARM family. */
53 
54 /* define either of these (preferably FLOAT_LOOKUP) to have faster
55    but less precise implementation. */
56 #undef FLOAT_LOOKUP
57 #undef INT_LOOKUP
58 
59 #ifdef FLOAT_LOOKUP
60 #include "lookup.c" /* catch this in the build system; we #include for
61                        compilers (like gcc) that can't inline across
62                        modules */
63 
64 /* side effect: changes *lsp to cosines of lsp */
vorbis_lsp_to_curve(float * curve,int * map,int n,int ln,float * lsp,int m,float amp,float ampoffset)65 void vorbis_lsp_to_curve(float *curve,int *map,int n,int ln,float *lsp,int m,
66                             float amp,float ampoffset){
67   int i;
68   float wdel=M_PI/ln;
69   vorbis_fpu_control fpu;
70 
71   vorbis_fpu_setround(&fpu);
72   for(i=0;i<m;i++)lsp[i]=vorbis_coslook(lsp[i]);
73 
74   i=0;
75   while(i<n){
76     int k=map[i];
77     int qexp;
78     float p=.7071067812f;
79     float q=.7071067812f;
80     float w=vorbis_coslook(wdel*k);
81     float *ftmp=lsp;
82     int c=m>>1;
83 
84     do{
85       q*=ftmp[0]-w;
86       p*=ftmp[1]-w;
87       ftmp+=2;
88     }while(--c);
89 
90     if(m&1){
91       /* odd order filter; slightly assymetric */
92       /* the last coefficient */
93       q*=ftmp[0]-w;
94       q*=q;
95       p*=p*(1.f-w*w);
96     }else{
97       /* even order filter; still symmetric */
98       q*=q*(1.f+w);
99       p*=p*(1.f-w);
100     }
101 
102     q=frexp(p+q,&qexp);
103     q=vorbis_fromdBlook(amp*
104                         vorbis_invsqlook(q)*
105                         vorbis_invsq2explook(qexp+m)-
106                         ampoffset);
107 
108     do{
109       curve[i++]*=q;
110     }while(map[i]==k);
111   }
112   vorbis_fpu_restore(fpu);
113 }
114 
115 #else
116 
117 #ifdef INT_LOOKUP
118 #include "lookup.c" /* catch this in the build system; we #include for
119                        compilers (like gcc) that can't inline across
120                        modules */
121 
122 static const int MLOOP_1[64]={
123    0,10,11,11, 12,12,12,12, 13,13,13,13, 13,13,13,13,
124   14,14,14,14, 14,14,14,14, 14,14,14,14, 14,14,14,14,
125   15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15,
126   15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15,
127 };
128 
129 static const int MLOOP_2[64]={
130   0,4,5,5, 6,6,6,6, 7,7,7,7, 7,7,7,7,
131   8,8,8,8, 8,8,8,8, 8,8,8,8, 8,8,8,8,
132   9,9,9,9, 9,9,9,9, 9,9,9,9, 9,9,9,9,
133   9,9,9,9, 9,9,9,9, 9,9,9,9, 9,9,9,9,
134 };
135 
136 static const int MLOOP_3[8]={0,1,2,2,3,3,3,3};
137 
138 
139 /* side effect: changes *lsp to cosines of lsp */
vorbis_lsp_to_curve(float * curve,int * map,int n,int ln,float * lsp,int m,float amp,float ampoffset)140 void vorbis_lsp_to_curve(float *curve,int *map,int n,int ln,float *lsp,int m,
141                             float amp,float ampoffset){
142 
143   /* 0 <= m < 256 */
144 
145   /* set up for using all int later */
146   int i;
147   int ampoffseti=rint(ampoffset*4096.f);
148   int ampi=rint(amp*16.f);
149   long *ilsp=alloca(m*sizeof(*ilsp));
150   for(i=0;i<m;i++)ilsp[i]=vorbis_coslook_i(lsp[i]/M_PI*65536.f+.5f);
151 
152   i=0;
153   while(i<n){
154     int j,k=map[i];
155     unsigned long pi=46341; /* 2**-.5 in 0.16 */
156     unsigned long qi=46341;
157     int qexp=0,shift;
158     long wi=vorbis_coslook_i(k*65536/ln);
159 
160     qi*=labs(ilsp[0]-wi);
161     pi*=labs(ilsp[1]-wi);
162 
163     for(j=3;j<m;j+=2){
164       if(!(shift=MLOOP_1[(pi|qi)>>25]))
165         if(!(shift=MLOOP_2[(pi|qi)>>19]))
166           shift=MLOOP_3[(pi|qi)>>16];
167       qi=(qi>>shift)*labs(ilsp[j-1]-wi);
168       pi=(pi>>shift)*labs(ilsp[j]-wi);
169       qexp+=shift;
170     }
171     if(!(shift=MLOOP_1[(pi|qi)>>25]))
172       if(!(shift=MLOOP_2[(pi|qi)>>19]))
173         shift=MLOOP_3[(pi|qi)>>16];
174 
175     /* pi,qi normalized collectively, both tracked using qexp */
176 
177     if(m&1){
178       /* odd order filter; slightly assymetric */
179       /* the last coefficient */
180       qi=(qi>>shift)*labs(ilsp[j-1]-wi);
181       pi=(pi>>shift)<<14;
182       qexp+=shift;
183 
184       if(!(shift=MLOOP_1[(pi|qi)>>25]))
185         if(!(shift=MLOOP_2[(pi|qi)>>19]))
186           shift=MLOOP_3[(pi|qi)>>16];
187 
188       pi>>=shift;
189       qi>>=shift;
190       qexp+=shift-14*((m+1)>>1);
191 
192       pi=((pi*pi)>>16);
193       qi=((qi*qi)>>16);
194       qexp=qexp*2+m;
195 
196       pi*=(1<<14)-((wi*wi)>>14);
197       qi+=pi>>14;
198 
199     }else{
200       /* even order filter; still symmetric */
201 
202       /* p*=p(1-w), q*=q(1+w), let normalization drift because it isn't
203          worth tracking step by step */
204 
205       pi>>=shift;
206       qi>>=shift;
207       qexp+=shift-7*m;
208 
209       pi=((pi*pi)>>16);
210       qi=((qi*qi)>>16);
211       qexp=qexp*2+m;
212 
213       pi*=(1<<14)-wi;
214       qi*=(1<<14)+wi;
215       qi=(qi+pi)>>14;
216 
217     }
218 
219 
220     /* we've let the normalization drift because it wasn't important;
221        however, for the lookup, things must be normalized again.  We
222        need at most one right shift or a number of left shifts */
223 
224     if(qi&0xffff0000){ /* checks for 1.xxxxxxxxxxxxxxxx */
225       qi>>=1; qexp++;
226     }else
227       while(qi && !(qi&0x8000)){ /* checks for 0.0xxxxxxxxxxxxxxx or less*/
228         qi<<=1; qexp--;
229       }
230 
231     amp=vorbis_fromdBlook_i(ampi*                     /*  n.4         */
232                             vorbis_invsqlook_i(qi,qexp)-
233                                                       /*  m.8, m+n<=8 */
234                             ampoffseti);              /*  8.12[0]     */
235 
236     curve[i]*=amp;
237     while(map[++i]==k)curve[i]*=amp;
238   }
239 }
240 
241 #else
242 
243 /* old, nonoptimized but simple version for any poor sap who needs to
244    figure out what the hell this code does, or wants the other
245    fraction of a dB precision */
246 
247 /* side effect: changes *lsp to cosines of lsp */
vorbis_lsp_to_curve(float * curve,int * map,int n,int ln,float * lsp,int m,float amp,float ampoffset)248 void vorbis_lsp_to_curve(float *curve,int *map,int n,int ln,float *lsp,int m,
249                             float amp,float ampoffset){
250   int i;
251   float wdel=M_PI/ln;
252   for(i=0;i<m;i++)lsp[i]=2.f*cos(lsp[i]);
253 
254   i=0;
255   while(i<n){
256     int j,k=map[i];
257     float p=.5f;
258     float q=.5f;
259     float w=2.f*cos(wdel*k);
260     for(j=1;j<m;j+=2){
261       q *= w-lsp[j-1];
262       p *= w-lsp[j];
263     }
264     if(j==m){
265       /* odd order filter; slightly assymetric */
266       /* the last coefficient */
267       q*=w-lsp[j-1];
268       p*=p*(4.f-w*w);
269       q*=q;
270     }else{
271       /* even order filter; still symmetric */
272       p*=p*(2.f-w);
273       q*=q*(2.f+w);
274     }
275 
276     q=fromdB(amp/sqrt(p+q)-ampoffset);
277 
278     curve[i]*=q;
279     while(map[++i]==k)curve[i]*=q;
280   }
281 }
282 
283 #endif
284 #endif
285 
cheby(float * g,int ord)286 static void cheby(float *g, int ord) {
287   int i, j;
288 
289   g[0] *= .5f;
290   for(i=2; i<= ord; i++) {
291     for(j=ord; j >= i; j--) {
292       g[j-2] -= g[j];
293       g[j] += g[j];
294     }
295   }
296 }
297 
comp(const void * a,const void * b)298 static int comp(const void *a,const void *b){
299   return (*(float *)a<*(float *)b)-(*(float *)a>*(float *)b);
300 }
301 
302 /* Newton-Raphson-Maehly actually functioned as a decent root finder,
303    but there are root sets for which it gets into limit cycles
304    (exacerbated by zero suppression) and fails.  We can't afford to
305    fail, even if the failure is 1 in 100,000,000, so we now use
306    Laguerre and later polish with Newton-Raphson (which can then
307    afford to fail) */
308 
309 #define EPSILON 10e-7
Laguerre_With_Deflation(float * a,int ord,float * r)310 static int Laguerre_With_Deflation(float *a,int ord,float *r){
311   int i,m;
312   double lastdelta=0.f;
313   double *defl=alloca(sizeof(*defl)*(ord+1));
314   for(i=0;i<=ord;i++)defl[i]=a[i];
315 
316   for(m=ord;m>0;m--){
317     double new=0.f,delta;
318 
319     /* iterate a root */
320     while(1){
321       double p=defl[m],pp=0.f,ppp=0.f,denom;
322 
323       /* eval the polynomial and its first two derivatives */
324       for(i=m;i>0;i--){
325         ppp = new*ppp + pp;
326         pp  = new*pp  + p;
327         p   = new*p   + defl[i-1];
328       }
329 
330       /* Laguerre's method */
331       denom=(m-1) * ((m-1)*pp*pp - m*p*ppp);
332       if(denom<0)
333         return(-1);  /* complex root!  The LPC generator handed us a bad filter */
334 
335       if(pp>0){
336         denom = pp + sqrt(denom);
337         if(denom<EPSILON)denom=EPSILON;
338       }else{
339         denom = pp - sqrt(denom);
340         if(denom>-(EPSILON))denom=-(EPSILON);
341       }
342 
343       delta  = m*p/denom;
344       new   -= delta;
345 
346       if(delta<0.f)delta*=-1;
347 
348       if(fabs(delta/new)<10e-12)break;
349       lastdelta=delta;
350     }
351 
352     r[m-1]=new;
353 
354     /* forward deflation */
355 
356     for(i=m;i>0;i--)
357       defl[i-1]+=new*defl[i];
358     defl++;
359 
360   }
361   return(0);
362 }
363 
364 
365 /* for spit-and-polish only */
Newton_Raphson(float * a,int ord,float * r)366 static int Newton_Raphson(float *a,int ord,float *r){
367   int i, k, count=0;
368   double error=1.f;
369   double *root=alloca(ord*sizeof(*root));
370 
371   for(i=0; i<ord;i++) root[i] = r[i];
372 
373   while(error>1e-20){
374     error=0;
375 
376     for(i=0; i<ord; i++) { /* Update each point. */
377       double pp=0.,delta;
378       double rooti=root[i];
379       double p=a[ord];
380       for(k=ord-1; k>= 0; k--) {
381 
382         pp= pp* rooti + p;
383         p = p * rooti + a[k];
384       }
385 
386       delta = p/pp;
387       root[i] -= delta;
388       error+= delta*delta;
389     }
390 
391     if(count>40)return(-1);
392 
393     count++;
394   }
395 
396   /* Replaced the original bubble sort with a real sort.  With your
397      help, we can eliminate the bubble sort in our lifetime. --Monty */
398 
399   for(i=0; i<ord;i++) r[i] = root[i];
400   return(0);
401 }
402 
403 
404 /* Convert lpc coefficients to lsp coefficients */
vorbis_lpc_to_lsp(float * lpc,float * lsp,int m)405 int vorbis_lpc_to_lsp(float *lpc,float *lsp,int m){
406   int order2=(m+1)>>1;
407   int g1_order,g2_order;
408   float *g1=alloca(sizeof(*g1)*(order2+1));
409   float *g2=alloca(sizeof(*g2)*(order2+1));
410   float *g1r=alloca(sizeof(*g1r)*(order2+1));
411   float *g2r=alloca(sizeof(*g2r)*(order2+1));
412   int i;
413 
414   /* even and odd are slightly different base cases */
415   g1_order=(m+1)>>1;
416   g2_order=(m)  >>1;
417 
418   /* Compute the lengths of the x polynomials. */
419   /* Compute the first half of K & R F1 & F2 polynomials. */
420   /* Compute half of the symmetric and antisymmetric polynomials. */
421   /* Remove the roots at +1 and -1. */
422 
423   g1[g1_order] = 1.f;
424   for(i=1;i<=g1_order;i++) g1[g1_order-i] = lpc[i-1]+lpc[m-i];
425   g2[g2_order] = 1.f;
426   for(i=1;i<=g2_order;i++) g2[g2_order-i] = lpc[i-1]-lpc[m-i];
427 
428   if(g1_order>g2_order){
429     for(i=2; i<=g2_order;i++) g2[g2_order-i] += g2[g2_order-i+2];
430   }else{
431     for(i=1; i<=g1_order;i++) g1[g1_order-i] -= g1[g1_order-i+1];
432     for(i=1; i<=g2_order;i++) g2[g2_order-i] += g2[g2_order-i+1];
433   }
434 
435   /* Convert into polynomials in cos(alpha) */
436   cheby(g1,g1_order);
437   cheby(g2,g2_order);
438 
439   /* Find the roots of the 2 even polynomials.*/
440   if(Laguerre_With_Deflation(g1,g1_order,g1r) ||
441      Laguerre_With_Deflation(g2,g2_order,g2r))
442     return(-1);
443 
444   Newton_Raphson(g1,g1_order,g1r); /* if it fails, it leaves g1r alone */
445   Newton_Raphson(g2,g2_order,g2r); /* if it fails, it leaves g2r alone */
446 
447   qsort(g1r,g1_order,sizeof(*g1r),comp);
448   qsort(g2r,g2_order,sizeof(*g2r),comp);
449 
450   for(i=0;i<g1_order;i++)
451     lsp[i*2] = acos(g1r[i]);
452 
453   for(i=0;i<g2_order;i++)
454     lsp[i*2+1] = acos(g2r[i]);
455   return(0);
456 }
457