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 1.12 2000/11/06 00:07:01 xiphmont Exp $
17 The LSP generation code is taken (with minimal modification) from
18 "On the Computation of the LSP Frequencies" by Joseph Rothweiler
19 <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);
89 q=frexp(p*p*(1.+w)+q*q*(1.-w),&qexp);
90 q=vorbis_fromdBlook(amp*
92 vorbis_invsq2explook(qexp+m)-
99 vorbis_fpu_restore(fpu);
105 #include "lookup.c" /* catch this in the build system; we #include for
106 compilers (like gcc) that can't inline across
109 static int MLOOP_1[64]={
110 0,10,11,11, 12,12,12,12, 13,13,13,13, 13,13,13,13,
111 14,14,14,14, 14,14,14,14, 14,14,14,14, 14,14,14,14,
112 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15,
113 15,15,15,15, 15,15,15,15, 15,15,15,15, 15,15,15,15,
116 static int MLOOP_2[64]={
117 0,4,5,5, 6,6,6,6, 7,7,7,7, 7,7,7,7,
118 8,8,8,8, 8,8,8,8, 8,8,8,8, 8,8,8,8,
119 9,9,9,9, 9,9,9,9, 9,9,9,9, 9,9,9,9,
120 9,9,9,9, 9,9,9,9, 9,9,9,9, 9,9,9,9,
123 static int MLOOP_3[8]={0,1,2,2,3,3,3,3};
126 /* side effect: changes *lsp to cosines of lsp */
127 void vorbis_lsp_to_curve(float *curve,int *map,int n,int ln,float *lsp,int m,
128 float amp,float ampoffset){
132 /* set up for using all int later */
134 int ampoffseti=rint(ampoffset*4096.);
135 int ampi=rint(amp*16.);
136 long *ilsp=alloca(m*sizeof(long));
137 for(i=0;i<m;i++)ilsp[i]=vorbis_coslook_i(lsp[i]/M_PI*65536.+.5);
142 unsigned long pi=46341; /* 2**-.5 in 0.16 */
143 unsigned long qi=46341;
145 long wi=vorbis_coslook_i(k*65536/ln);
147 pi*=labs(ilsp[0]-wi);
148 qi*=labs(ilsp[1]-wi);
151 if(!(shift=MLOOP_1[(pi|qi)>>25]))
152 if(!(shift=MLOOP_2[(pi|qi)>>19]))
153 shift=MLOOP_3[(pi|qi)>>16];
154 pi=(pi>>shift)*labs(ilsp[j]-wi);
155 qi=(qi>>shift)*labs(ilsp[j+1]-wi);
158 if(!(shift=MLOOP_1[(pi|qi)>>25]))
159 if(!(shift=MLOOP_2[(pi|qi)>>19]))
160 shift=MLOOP_3[(pi|qi)>>16];
165 /* pi,qi normalized collectively, both tracked using qexp */
167 /* p*=p(1-w), q*=q(1+w), let normalization drift because it isn't
168 worth tracking step by step */
179 /* we've let the normalization drift because it wasn't important;
180 however, for the lookup, things must be normalized again. We
181 need at most one right shift or a number of left shifts */
183 if(qi&0xffff0000){ /* checks for 1.xxxxxxxxxxxxxxxx */
186 while(qi && !(qi&0x8000)){ /* checks for 0.0xxxxxxxxxxxxxxx or less*/
190 amp=vorbis_fromdBlook_i(ampi* /* n.4 */
191 vorbis_invsqlook_i(qi,qexp)-
193 ampoffseti); /* 8.12[0] */
196 while(map[++i]==k)curve[i]=amp;
202 /* old, nonoptimized but simple version for any poor sap who needs to
203 figure out what the hell this code does, or wants the other tiny
204 fraction of a dB precision */
206 /* side effect: changes *lsp to cosines of lsp */
207 void vorbis_lsp_to_curve(float *curve,int *map,int n,int ln,float *lsp,int m,
208 float amp,float ampoffset){
211 for(i=0;i<m;i++)lsp[i]=2*cos(lsp[i]);
218 float w=2*cos(wdel*k);
225 q=fromdB(amp/sqrt(p+q)-ampoffset);
228 while(map[++i]==k)curve[i]=q;
235 static void cheby(float *g, int ord) {
239 for(i=2; i<= ord; i++) {
240 for(j=ord; j >= i; j--) {
247 static int comp(const void *a,const void *b){
248 if(*(float *)a<*(float *)b)
254 /* This is one of those 'mathemeticians should not write code' kind of
255 cases. Newton's method of polishing roots is straightforward
256 enough... except in those cases where it just fails in the real
257 world. In our case below, we're worried about a local mini/maxima
258 shooting a root estimation off to infinity, or the new estimation
259 chaotically oscillating about convergence (shouldn't actually be a
260 problem in our usage.
262 Maehly's modification (zero suppression, to prevent two tenative
263 roots from collapsing to the same actual root) similarly can
264 temporarily shoot a root off toward infinity. It would come
265 back... if it were not for the fact that machine representation has
266 limited dynamic range and resolution. This too is guarded by
269 Last problem is convergence criteria; we don't know what a 'double'
270 is on our hardware/compiler, and the convergence limit is bounded
271 by roundoff noise. So, we hack convergence:
273 Require at most 1e-6 mean squared error for all zeroes. When
274 converging, start the clock ticking at 1e-6; limit our polishing to
275 as many more iterations as took us to get this far, 100 max.
277 Past max iters, quit when MSE is no longer decreasing *or* we go
278 below ~1e-20 MSE, whichever happens first. */
280 static void Newton_Raphson_Maehly(float *a,int ord,float *r){
281 int i, k, count=0, maxiter=0;
282 double error=1.,besterror=1.;
283 double *root=alloca(ord*sizeof(double));
285 for(i=0; i<ord;i++) root[i] = 2.0 * (i+0.5) / ord - 1.0;
290 for(i=0; i<ord; i++) { /* Update each point. */
291 double ac=0.,pp=0.,delta;
292 double rooti=root[i];
294 for(k=ord-1; k>= 0; k--) {
298 if (k != i) ac += 1./(rooti - root[k]);
304 /* don't allow the correction to scream off into infinity if we
305 happened to polish right at a local mini/maximum */
307 if(delta<-3)delta=-3;
308 if(delta>3.)delta=3.; /* 3 is not a random choice; it's large
309 enough to make sure the first pass
310 can't accidentally limit two poles to
311 the same value in a fatal nonelastic
315 error += delta*delta;
318 if(maxiter && count>maxiter && error>=besterror)break;
320 /* anything to help out the polisher; converge using doubles */
321 if(!count || error<besterror){
322 for(i=0; i<ord; i++) r[i]=root[i];
324 if(error<1.e-6){ /* rough minimum criteria */
326 if(maxiter>100)maxiter=100;
333 /* Replaced the original bubble sort with a real sort. With your
334 help, we can eliminate the bubble sort in our lifetime. --Monty */
336 qsort(r,ord,sizeof(float),comp);
340 /* Convert lpc coefficients to lsp coefficients */
341 void vorbis_lpc_to_lsp(float *lpc,float *lsp,int m){
343 float *g1=alloca(sizeof(float)*(order2+1));
344 float *g2=alloca(sizeof(float)*(order2+1));
345 float *g1r=alloca(sizeof(float)*(order2+1));
346 float *g2r=alloca(sizeof(float)*(order2+1));
349 /* Compute the lengths of the x polynomials. */
350 /* Compute the first half of K & R F1 & F2 polynomials. */
351 /* Compute half of the symmetric and antisymmetric polynomials. */
352 /* Remove the roots at +1 and -1. */
355 for(i=0;i<order2;i++) g1[order2-i-1] = lpc[i]+lpc[m-i-1];
357 for(i=0;i<order2;i++) g2[order2-i-1] = lpc[i]-lpc[m-i-1];
359 for(i=0; i<order2;i++) g1[order2-i-1] -= g1[order2-i];
360 for(i=0; i<order2;i++) g2[order2-i-1] += g2[order2-i];
362 /* Convert into polynomials in cos(alpha) */
366 /* Find the roots of the 2 even polynomials.*/
368 Newton_Raphson_Maehly(g1,order2,g1r);
369 Newton_Raphson_Maehly(g2,order2,g2r);
372 lsp[i] = acos(g1r[i/2]);
373 lsp[i+1] = acos(g2r[i/2]);