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 1994-2000 *
9 * by Monty <monty@xiph.org> and the XIPHOPHORUS Company *
10 * http://www.xiph.org/ *
12 ********************************************************************
14 function: LSP (also called LSF) conversion routines
15 last mod: $Id: lsp.c,v 1.14 2001/01/22 01:38:25 xiphmont Exp $
17 The LSP generation code is taken (with minimal modification and a
18 few bugfixes) from "On the Computation of the LSP Frequencies" by
19 Joseph Rothweiler <rothwlr@altavista.net>, available at:
21 http://www2.xtdl.com/~rothwlr/lsfpaper/lsfpage.html
23 ********************************************************************/
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
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
54 /* undefine both for the 'old' but more precise implementation */
59 #include "lookup.c" /* catch this in the build system; we #include for
60 compilers (like gcc) that can't inline across
63 /* side effect: changes *lsp to cosines of lsp */
64 void vorbis_lsp_to_curve(float *curve,int *map,int n,int ln,float *lsp,int m,
65 float amp,float ampoffset){
68 vorbis_fpu_control fpu;
70 vorbis_fpu_setround(&fpu);
71 for(i=0;i<m;i++)lsp[i]=vorbis_coslook(lsp[i]);
79 float w=vorbis_coslook(wdel*k);
90 /* odd order filter; slightly assymetric */
91 /* the last coefficient */
96 /* even order filter; still symmetric */
102 q=vorbis_fromdBlook(amp*
104 vorbis_invsq2explook(qexp+m)-
111 vorbis_fpu_restore(fpu);
117 #include "lookup.c" /* catch this in the build system; we #include for
118 compilers (like gcc) that can't inline across
121 static int MLOOP_1[64]={
122 0,10,11,11, 12,12,12,12, 13,13,13,13, 13,13,13,13,
123 14,14,14,14, 14,14,14,14, 14,14,14,14, 14,14,14,14,
124 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15,
125 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15,
128 static int MLOOP_2[64]={
129 0,4,5,5, 6,6,6,6, 7,7,7,7, 7,7,7,7,
130 8,8,8,8, 8,8,8,8, 8,8,8,8, 8,8,8,8,
131 9,9,9,9, 9,9,9,9, 9,9,9,9, 9,9,9,9,
132 9,9,9,9, 9,9,9,9, 9,9,9,9, 9,9,9,9,
135 static int MLOOP_3[8]={0,1,2,2,3,3,3,3};
138 /* side effect: changes *lsp to cosines of lsp */
139 void vorbis_lsp_to_curve(float *curve,int *map,int n,int ln,float *lsp,int m,
140 float amp,float ampoffset){
144 /* set up for using all int later */
146 int ampoffseti=rint(ampoffset*4096.f);
147 int ampi=rint(amp*16.f);
148 long *ilsp=alloca(m*sizeof(long));
149 for(i=0;i<m;i++)ilsp[i]=vorbis_coslook_i(lsp[i]/M_PI*65536.f+.5f);
154 unsigned long pi=46341; /* 2**-.5 in 0.16 */
155 unsigned long qi=46341;
157 long wi=vorbis_coslook_i(k*65536/ln);
159 qi*=labs(ilsp[0]-wi);
160 pi*=labs(ilsp[1]-wi);
163 if(!(shift=MLOOP_1[(pi|qi)>>25]))
164 if(!(shift=MLOOP_2[(pi|qi)>>19]))
165 shift=MLOOP_3[(pi|qi)>>16];
166 qi=(qi>>shift)*labs(ilsp[j-1]-wi);
167 pi=(pi>>shift)*labs(ilsp[j]-wi);
170 if(!(shift=MLOOP_1[(pi|qi)>>25]))
171 if(!(shift=MLOOP_2[(pi|qi)>>19]))
172 shift=MLOOP_3[(pi|qi)>>16];
174 /* pi,qi normalized collectively, both tracked using qexp */
177 /* odd order filter; slightly assymetric */
178 /* the last coefficient */
179 qi=(qi>>shift)*labs(ilsp[j-1]-wi);
183 if(!(shift=MLOOP_1[(pi|qi)>>25]))
184 if(!(shift=MLOOP_2[(pi|qi)>>19]))
185 shift=MLOOP_3[(pi|qi)>>16];
189 qexp+=shift-14*((m+1)>>1);
195 pi*=(1<<14)-((wi*wi)>>14);
202 /* even order filter; still symmetric */
204 /* p*=p(1-w), q*=q(1+w), let normalization drift because it isn't
205 worth tracking step by step */
222 /* we've let the normalization drift because it wasn't important;
223 however, for the lookup, things must be normalized again. We
224 need at most one right shift or a number of left shifts */
226 if(qi&0xffff0000){ /* checks for 1.xxxxxxxxxxxxxxxx */
229 while(qi && !(qi&0x8000)){ /* checks for 0.0xxxxxxxxxxxxxxx or less*/
233 amp=vorbis_fromdBlook_i(ampi* /* n.4 */
234 vorbis_invsqlook_i(qi,qexp)-
236 ampoffseti); /* 8.12[0] */
239 while(map[++i]==k)curve[i]=amp;
245 /* old, nonoptimized but simple version for any poor sap who needs to
246 figure out what the hell this code does, or wants the other
247 fraction of a dB precision */
249 /* side effect: changes *lsp to cosines of lsp */
250 void vorbis_lsp_to_curve(float *curve,int *map,int n,int ln,float *lsp,int m,
251 float amp,float ampoffset){
254 for(i=0;i<m;i++)lsp[i]=2.f*cos(lsp[i]);
261 float w=2.f*cos(wdel*k);
267 /* odd order filter; slightly assymetric */
268 /* the last coefficient */
273 /* even order filter; still symmetric */
278 q=fromdB(amp/sqrt(p+q)-ampoffset);
281 while(map[++i]==k)curve[i]=q;
288 static void cheby(float *g, int ord) {
292 for(i=2; i<= ord; i++) {
293 for(j=ord; j >= i; j--) {
300 static int comp(const void *a,const void *b){
301 if(*(float *)a<*(float *)b)
307 /* This is one of those 'mathemeticians should not write code' kind of
308 cases. Newton's method of polishing roots is straightforward
309 enough... except in those cases where it just fails in the real
310 world. In our case below, we're worried about a local mini/maxima
311 shooting a root estimation off to infinity, or the new estimation
312 chaotically oscillating about convergence (shouldn't actually be a
313 problem in our usage.
315 Maehly's modification (zero suppression, to prevent two tenative
316 roots from collapsing to the same actual root) similarly can
317 temporarily shoot a root off toward infinity. It would come
318 back... if it were not for the fact that machine representation has
319 limited dynamic range and resolution. This too is guarded by
322 Last problem is convergence criteria; we don't know what a 'double'
323 is on our hardware/compiler, and the convergence limit is bounded
324 by roundoff noise. So, we hack convergence:
326 Require at most 1e-6 mean squared error for all zeroes. When
327 converging, start the clock ticking at 1e-6; limit our polishing to
328 as many more iterations as took us to get this far, 100 max.
330 Past max iters, quit when MSE is no longer decreasing *or* we go
331 below ~1e-20 MSE, whichever happens first. */
333 static void Newton_Raphson_Maehly(float *a,int ord,float *r){
334 int i, k, count=0, maxiter=0;
335 double error=1.,besterror=1.;
336 double *root=alloca(ord*sizeof(double));
338 for(i=0; i<ord;i++) root[i] = 2.0 * (i+0.5) / ord - 1.0;
343 for(i=0; i<ord; i++) { /* Update each point. */
344 double ac=0.,pp=0.,delta;
345 double rooti=root[i];
347 for(k=ord-1; k>= 0; k--) {
351 if (k != i) ac += 1./(rooti - root[k]);
357 /* don't allow the correction to scream off into infinity if we
358 happened to polish right at a local mini/maximum */
360 if(delta<-3.)delta=-3.;
361 if(delta>3.)delta=3.; /* 3 is not a random choice; it's large
362 enough to make sure the first pass
363 can't accidentally limit two poles to
364 the same value in a fatal nonelastic
368 error += delta*delta;
371 if(maxiter && count>maxiter && error>=besterror)break;
373 /* anything to help out the polisher; converge using doubles */
374 if(!count || error<besterror){
375 for(i=0; i<ord; i++) r[i]=root[i];
377 if(error<1e-6){ /* rough minimum criteria */
379 if(maxiter>100)maxiter=100;
386 /* Replaced the original bubble sort with a real sort. With your
387 help, we can eliminate the bubble sort in our lifetime. --Monty */
389 qsort(r,ord,sizeof(float),comp);
393 /* Convert lpc coefficients to lsp coefficients */
394 void vorbis_lpc_to_lsp(float *lpc,float *lsp,int m){
396 int g1_order,g2_order;
397 float *g1=alloca(sizeof(float)*(order2+1));
398 float *g2=alloca(sizeof(float)*(order2+1));
399 float *g1r=alloca(sizeof(float)*(order2+1));
400 float *g2r=alloca(sizeof(float)*(order2+1));
403 /* even and odd are slightly different base cases */
407 /* Compute the lengths of the x polynomials. */
408 /* Compute the first half of K & R F1 & F2 polynomials. */
409 /* Compute half of the symmetric and antisymmetric polynomials. */
410 /* Remove the roots at +1 and -1. */
413 for(i=1;i<=g1_order;i++) g1[g1_order-i] = lpc[i-1]+lpc[m-i];
415 for(i=1;i<=g2_order;i++) g2[g2_order-i] = lpc[i-1]-lpc[m-i];
417 if(g1_order>g2_order){
418 for(i=2; i<=g2_order;i++) g2[g2_order-i] += g2[g2_order-i+2];
420 for(i=1; i<=g1_order;i++) g1[g1_order-i] -= g1[g1_order-i+1];
421 for(i=1; i<=g2_order;i++) g2[g2_order-i] += g2[g2_order-i+1];
424 /* Convert into polynomials in cos(alpha) */
428 /* Find the roots of the 2 even polynomials.*/
430 Newton_Raphson_Maehly(g1,g1_order,g1r);
431 Newton_Raphson_Maehly(g2,g2_order,g2r);
433 for(i=0;i<g1_order;i++)
434 lsp[i*2] = acos(g1r[i]);
436 for(i=0;i<g2_order;i++)
437 lsp[i*2+1] = acos(g2r[i]);