1 /********************************************************************
3 * THIS FILE IS PART OF THE OggVorbis SOFTWARE CODEC SOURCE CODE. *
4 * USE, DISTRIBUTION AND REPRODUCTION OF THIS SOURCE IS GOVERNED BY *
5 * THE GNU LESSER/LIBRARY PUBLIC LICENSE, WHICH IS INCLUDED WITH *
6 * THIS SOURCE. PLEASE READ THESE TERMS BEFORE DISTRIBUTING. *
8 * THE OggVorbis SOURCE CODE IS (C) COPYRIGHT 1994-2001 *
9 * by the XIPHOPHORUS Company http://www.xiph.org/ *
11 ********************************************************************
13 function: LSP (also called LSF) conversion routines
14 last mod: $Id: lsp.c,v 1.15 2001/02/02 03:51:56 xiphmont Exp $
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 <rothwlr@altavista.net>, available at:
20 http://www2.xtdl.com/~rothwlr/lsfpaper/lsfpage.html
22 ********************************************************************/
24 /* Note that the lpc-lsp conversion finds the roots of polynomial with
25 an iterative root polisher (CACM algorithm 283). It *is* possible
26 to confuse this algorithm into not converging; that should only
27 happen with absurdly closely spaced roots (very sharp peaks in the
28 LPC f response) which in turn should be impossible in our use of
29 the code. If this *does* happen anyway, it's a bug in the floor
30 finder; find the cause of the confusion (probably a single bin
31 spike or accidental near-float-limit resolution problems) and
43 /* three possible LSP to f curve functions; the exact computation
44 (float), a lookup based float implementation, and an integer
45 implementation. The float lookup is likely the optimal choice on
46 any machine with an FPU. The integer implementation is *not* fixed
47 point (due to the need for a large dynamic range and thus a
48 seperately tracked exponent) and thus much more complex than the
49 relatively simple float implementations. It's mostly for future
50 work on a fully fixed point implementation for processors like the
53 /* undefine both for the 'old' but more precise implementation */
58 #include "lookup.c" /* catch this in the build system; we #include for
59 compilers (like gcc) that can't inline across
62 /* side effect: changes *lsp to cosines of lsp */
63 void vorbis_lsp_to_curve(float *curve,int *map,int n,int ln,float *lsp,int m,
64 float amp,float ampoffset){
67 vorbis_fpu_control fpu;
69 vorbis_fpu_setround(&fpu);
70 for(i=0;i<m;i++)lsp[i]=vorbis_coslook(lsp[i]);
78 float w=vorbis_coslook(wdel*k);
89 /* odd order filter; slightly assymetric */
90 /* the last coefficient */
95 /* even order filter; still symmetric */
101 q=vorbis_fromdBlook(amp*
103 vorbis_invsq2explook(qexp+m)-
110 vorbis_fpu_restore(fpu);
116 #include "lookup.c" /* catch this in the build system; we #include for
117 compilers (like gcc) that can't inline across
120 static int MLOOP_1[64]={
121 0,10,11,11, 12,12,12,12, 13,13,13,13, 13,13,13,13,
122 14,14,14,14, 14,14,14,14, 14,14,14,14, 14,14,14,14,
123 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15,
124 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15,
127 static int MLOOP_2[64]={
128 0,4,5,5, 6,6,6,6, 7,7,7,7, 7,7,7,7,
129 8,8,8,8, 8,8,8,8, 8,8,8,8, 8,8,8,8,
130 9,9,9,9, 9,9,9,9, 9,9,9,9, 9,9,9,9,
131 9,9,9,9, 9,9,9,9, 9,9,9,9, 9,9,9,9,
134 static int MLOOP_3[8]={0,1,2,2,3,3,3,3};
137 /* side effect: changes *lsp to cosines of lsp */
138 void vorbis_lsp_to_curve(float *curve,int *map,int n,int ln,float *lsp,int m,
139 float amp,float ampoffset){
143 /* set up for using all int later */
145 int ampoffseti=rint(ampoffset*4096.f);
146 int ampi=rint(amp*16.f);
147 long *ilsp=alloca(m*sizeof(long));
148 for(i=0;i<m;i++)ilsp[i]=vorbis_coslook_i(lsp[i]/M_PI*65536.f+.5f);
153 unsigned long pi=46341; /* 2**-.5 in 0.16 */
154 unsigned long qi=46341;
156 long wi=vorbis_coslook_i(k*65536/ln);
158 qi*=labs(ilsp[0]-wi);
159 pi*=labs(ilsp[1]-wi);
162 if(!(shift=MLOOP_1[(pi|qi)>>25]))
163 if(!(shift=MLOOP_2[(pi|qi)>>19]))
164 shift=MLOOP_3[(pi|qi)>>16];
165 qi=(qi>>shift)*labs(ilsp[j-1]-wi);
166 pi=(pi>>shift)*labs(ilsp[j]-wi);
169 if(!(shift=MLOOP_1[(pi|qi)>>25]))
170 if(!(shift=MLOOP_2[(pi|qi)>>19]))
171 shift=MLOOP_3[(pi|qi)>>16];
173 /* pi,qi normalized collectively, both tracked using qexp */
176 /* odd order filter; slightly assymetric */
177 /* the last coefficient */
178 qi=(qi>>shift)*labs(ilsp[j-1]-wi);
182 if(!(shift=MLOOP_1[(pi|qi)>>25]))
183 if(!(shift=MLOOP_2[(pi|qi)>>19]))
184 shift=MLOOP_3[(pi|qi)>>16];
188 qexp+=shift-14*((m+1)>>1);
194 pi*=(1<<14)-((wi*wi)>>14);
201 /* even order filter; still symmetric */
203 /* p*=p(1-w), q*=q(1+w), let normalization drift because it isn't
204 worth tracking step by step */
221 /* we've let the normalization drift because it wasn't important;
222 however, for the lookup, things must be normalized again. We
223 need at most one right shift or a number of left shifts */
225 if(qi&0xffff0000){ /* checks for 1.xxxxxxxxxxxxxxxx */
228 while(qi && !(qi&0x8000)){ /* checks for 0.0xxxxxxxxxxxxxxx or less*/
232 amp=vorbis_fromdBlook_i(ampi* /* n.4 */
233 vorbis_invsqlook_i(qi,qexp)-
235 ampoffseti); /* 8.12[0] */
238 while(map[++i]==k)curve[i]=amp;
244 /* old, nonoptimized but simple version for any poor sap who needs to
245 figure out what the hell this code does, or wants the other
246 fraction of a dB precision */
248 /* side effect: changes *lsp to cosines of lsp */
249 void vorbis_lsp_to_curve(float *curve,int *map,int n,int ln,float *lsp,int m,
250 float amp,float ampoffset){
253 for(i=0;i<m;i++)lsp[i]=2.f*cos(lsp[i]);
260 float w=2.f*cos(wdel*k);
266 /* odd order filter; slightly assymetric */
267 /* the last coefficient */
272 /* even order filter; still symmetric */
277 q=fromdB(amp/sqrt(p+q)-ampoffset);
280 while(map[++i]==k)curve[i]=q;
287 static void cheby(float *g, int ord) {
291 for(i=2; i<= ord; i++) {
292 for(j=ord; j >= i; j--) {
299 static int comp(const void *a,const void *b){
300 if(*(float *)a<*(float *)b)
306 /* This is one of those 'mathemeticians should not write code' kind of
307 cases. Newton's method of polishing roots is straightforward
308 enough... except in those cases where it just fails in the real
309 world. In our case below, we're worried about a local mini/maxima
310 shooting a root estimation off to infinity, or the new estimation
311 chaotically oscillating about convergence (shouldn't actually be a
312 problem in our usage.
314 Maehly's modification (zero suppression, to prevent two tenative
315 roots from collapsing to the same actual root) similarly can
316 temporarily shoot a root off toward infinity. It would come
317 back... if it were not for the fact that machine representation has
318 limited dynamic range and resolution. This too is guarded by
321 Last problem is convergence criteria; we don't know what a 'double'
322 is on our hardware/compiler, and the convergence limit is bounded
323 by roundoff noise. So, we hack convergence:
325 Require at most 1e-6 mean squared error for all zeroes. When
326 converging, start the clock ticking at 1e-6; limit our polishing to
327 as many more iterations as took us to get this far, 100 max.
329 Past max iters, quit when MSE is no longer decreasing *or* we go
330 below ~1e-20 MSE, whichever happens first. */
332 static void Newton_Raphson_Maehly(float *a,int ord,float *r){
333 int i, k, count=0, maxiter=0;
334 double error=1.,besterror=1.;
335 double *root=alloca(ord*sizeof(double));
337 for(i=0; i<ord;i++) root[i] = 2.0 * (i+0.5) / ord - 1.0;
342 for(i=0; i<ord; i++) { /* Update each point. */
343 double ac=0.,pp=0.,delta;
344 double rooti=root[i];
346 for(k=ord-1; k>= 0; k--) {
350 if (k != i) ac += 1./(rooti - root[k]);
356 /* don't allow the correction to scream off into infinity if we
357 happened to polish right at a local mini/maximum */
359 if(delta<-3.)delta=-3.;
360 if(delta>3.)delta=3.; /* 3 is not a random choice; it's large
361 enough to make sure the first pass
362 can't accidentally limit two poles to
363 the same value in a fatal nonelastic
367 error += delta*delta;
370 if(maxiter && count>maxiter && error>=besterror)break;
372 /* anything to help out the polisher; converge using doubles */
373 if(!count || error<besterror){
374 for(i=0; i<ord; i++) r[i]=root[i];
376 if(error<1e-6){ /* rough minimum criteria */
378 if(maxiter>100)maxiter=100;
385 /* Replaced the original bubble sort with a real sort. With your
386 help, we can eliminate the bubble sort in our lifetime. --Monty */
388 qsort(r,ord,sizeof(float),comp);
392 /* Convert lpc coefficients to lsp coefficients */
393 void vorbis_lpc_to_lsp(float *lpc,float *lsp,int m){
395 int g1_order,g2_order;
396 float *g1=alloca(sizeof(float)*(order2+1));
397 float *g2=alloca(sizeof(float)*(order2+1));
398 float *g1r=alloca(sizeof(float)*(order2+1));
399 float *g2r=alloca(sizeof(float)*(order2+1));
402 /* even and odd are slightly different base cases */
406 /* Compute the lengths of the x polynomials. */
407 /* Compute the first half of K & R F1 & F2 polynomials. */
408 /* Compute half of the symmetric and antisymmetric polynomials. */
409 /* Remove the roots at +1 and -1. */
412 for(i=1;i<=g1_order;i++) g1[g1_order-i] = lpc[i-1]+lpc[m-i];
414 for(i=1;i<=g2_order;i++) g2[g2_order-i] = lpc[i-1]-lpc[m-i];
416 if(g1_order>g2_order){
417 for(i=2; i<=g2_order;i++) g2[g2_order-i] += g2[g2_order-i+2];
419 for(i=1; i<=g1_order;i++) g1[g1_order-i] -= g1[g1_order-i+1];
420 for(i=1; i<=g2_order;i++) g2[g2_order-i] += g2[g2_order-i+1];
423 /* Convert into polynomials in cos(alpha) */
427 /* Find the roots of the 2 even polynomials.*/
429 Newton_Raphson_Maehly(g1,g1_order,g1r);
430 Newton_Raphson_Maehly(g2,g2_order,g2r);
432 for(i=0;i<g1_order;i++)
433 lsp[i*2] = acos(g1r[i]);
435 for(i=0;i<g2_order;i++)
436 lsp[i*2+1] = acos(g2r[i]);