• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright (c) 2007-2008 CSIRO
2    Copyright (c) 2007-2009 Xiph.Org Foundation
3    Written by Jean-Marc Valin */
4 /**
5    @file pitch.c
6    @brief Pitch analysis
7  */
8 
9 /*
10    Redistribution and use in source and binary forms, with or without
11    modification, are permitted provided that the following conditions
12    are met:
13 
14    - Redistributions of source code must retain the above copyright
15    notice, this list of conditions and the following disclaimer.
16 
17    - Redistributions in binary form must reproduce the above copyright
18    notice, this list of conditions and the following disclaimer in the
19    documentation and/or other materials provided with the distribution.
20 
21    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22    ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23    LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24    A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
25    OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
26    EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
27    PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
28    PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
29    LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
30    NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
31    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32 */
33 
34 #ifdef HAVE_CONFIG_H
35 #include "config.h"
36 #endif
37 
38 #include "pitch.h"
39 #include "os_support.h"
40 #include "modes.h"
41 #include "stack_alloc.h"
42 #include "mathops.h"
43 #include "celt_lpc.h"
44 
find_best_pitch(opus_val32 * xcorr,opus_val16 * y,int len,int max_pitch,int * best_pitch,int yshift,opus_val32 maxcorr)45 static void find_best_pitch(opus_val32 *xcorr, opus_val16 *y, int len,
46                             int max_pitch, int *best_pitch
47 #ifdef FIXED_POINT
48                             , int yshift, opus_val32 maxcorr
49 #endif
50                             )
51 {
52    int i, j;
53    opus_val32 Syy=1;
54    opus_val16 best_num[2];
55    opus_val32 best_den[2];
56 #ifdef FIXED_POINT
57    int xshift;
58 
59    xshift = celt_ilog2(maxcorr)-14;
60 #endif
61 
62    best_num[0] = -1;
63    best_num[1] = -1;
64    best_den[0] = 0;
65    best_den[1] = 0;
66    best_pitch[0] = 0;
67    best_pitch[1] = 1;
68    for (j=0;j<len;j++)
69       Syy = ADD32(Syy, SHR32(MULT16_16(y[j],y[j]), yshift));
70    for (i=0;i<max_pitch;i++)
71    {
72       if (xcorr[i]>0)
73       {
74          opus_val16 num;
75          opus_val32 xcorr16;
76          xcorr16 = EXTRACT16(VSHR32(xcorr[i], xshift));
77 #ifndef FIXED_POINT
78          /* Considering the range of xcorr16, this should avoid both underflows
79             and overflows (inf) when squaring xcorr16 */
80          xcorr16 *= 1e-12f;
81 #endif
82          num = MULT16_16_Q15(xcorr16,xcorr16);
83          if (MULT16_32_Q15(num,best_den[1]) > MULT16_32_Q15(best_num[1],Syy))
84          {
85             if (MULT16_32_Q15(num,best_den[0]) > MULT16_32_Q15(best_num[0],Syy))
86             {
87                best_num[1] = best_num[0];
88                best_den[1] = best_den[0];
89                best_pitch[1] = best_pitch[0];
90                best_num[0] = num;
91                best_den[0] = Syy;
92                best_pitch[0] = i;
93             } else {
94                best_num[1] = num;
95                best_den[1] = Syy;
96                best_pitch[1] = i;
97             }
98          }
99       }
100       Syy += SHR32(MULT16_16(y[i+len],y[i+len]),yshift) - SHR32(MULT16_16(y[i],y[i]),yshift);
101       Syy = MAX32(1, Syy);
102    }
103 }
104 
celt_fir5(const opus_val16 * x,const opus_val16 * num,opus_val16 * y,int N,opus_val16 * mem)105 static void celt_fir5(const opus_val16 *x,
106          const opus_val16 *num,
107          opus_val16 *y,
108          int N,
109          opus_val16 *mem)
110 {
111    int i;
112    opus_val16 num0, num1, num2, num3, num4;
113    opus_val32 mem0, mem1, mem2, mem3, mem4;
114    num0=num[0];
115    num1=num[1];
116    num2=num[2];
117    num3=num[3];
118    num4=num[4];
119    mem0=mem[0];
120    mem1=mem[1];
121    mem2=mem[2];
122    mem3=mem[3];
123    mem4=mem[4];
124    for (i=0;i<N;i++)
125    {
126       opus_val32 sum = SHL32(EXTEND32(x[i]), SIG_SHIFT);
127       sum = MAC16_16(sum,num0,mem0);
128       sum = MAC16_16(sum,num1,mem1);
129       sum = MAC16_16(sum,num2,mem2);
130       sum = MAC16_16(sum,num3,mem3);
131       sum = MAC16_16(sum,num4,mem4);
132       mem4 = mem3;
133       mem3 = mem2;
134       mem2 = mem1;
135       mem1 = mem0;
136       mem0 = x[i];
137       y[i] = ROUND16(sum, SIG_SHIFT);
138    }
139    mem[0]=mem0;
140    mem[1]=mem1;
141    mem[2]=mem2;
142    mem[3]=mem3;
143    mem[4]=mem4;
144 }
145 
146 
pitch_downsample(celt_sig * OPUS_RESTRICT x[],opus_val16 * OPUS_RESTRICT x_lp,int len,int C,int arch)147 void pitch_downsample(celt_sig * OPUS_RESTRICT x[], opus_val16 * OPUS_RESTRICT x_lp,
148       int len, int C, int arch)
149 {
150    int i;
151    opus_val32 ac[5];
152    opus_val16 tmp=Q15ONE;
153    opus_val16 lpc[4], mem[5]={0,0,0,0,0};
154    opus_val16 lpc2[5];
155    opus_val16 c1 = QCONST16(.8f,15);
156 #ifdef FIXED_POINT
157    int shift;
158    opus_val32 maxabs = celt_maxabs32(x[0], len);
159    if (C==2)
160    {
161       opus_val32 maxabs_1 = celt_maxabs32(x[1], len);
162       maxabs = MAX32(maxabs, maxabs_1);
163    }
164    if (maxabs<1)
165       maxabs=1;
166    shift = celt_ilog2(maxabs)-10;
167    if (shift<0)
168       shift=0;
169    if (C==2)
170       shift++;
171 #endif
172    for (i=1;i<len>>1;i++)
173       x_lp[i] = SHR32(HALF32(HALF32(x[0][(2*i-1)]+x[0][(2*i+1)])+x[0][2*i]), shift);
174    x_lp[0] = SHR32(HALF32(HALF32(x[0][1])+x[0][0]), shift);
175    if (C==2)
176    {
177       for (i=1;i<len>>1;i++)
178          x_lp[i] += SHR32(HALF32(HALF32(x[1][(2*i-1)]+x[1][(2*i+1)])+x[1][2*i]), shift);
179       x_lp[0] += SHR32(HALF32(HALF32(x[1][1])+x[1][0]), shift);
180    }
181 
182    _celt_autocorr(x_lp, ac, NULL, 0,
183                   4, len>>1, arch);
184 
185    /* Noise floor -40 dB */
186 #ifdef FIXED_POINT
187    ac[0] += SHR32(ac[0],13);
188 #else
189    ac[0] *= 1.0001f;
190 #endif
191    /* Lag windowing */
192    for (i=1;i<=4;i++)
193    {
194       /*ac[i] *= exp(-.5*(2*M_PI*.002*i)*(2*M_PI*.002*i));*/
195 #ifdef FIXED_POINT
196       ac[i] -= MULT16_32_Q15(2*i*i, ac[i]);
197 #else
198       ac[i] -= ac[i]*(.008f*i)*(.008f*i);
199 #endif
200    }
201 
202    _celt_lpc(lpc, ac, 4);
203    for (i=0;i<4;i++)
204    {
205       tmp = MULT16_16_Q15(QCONST16(.9f,15), tmp);
206       lpc[i] = MULT16_16_Q15(lpc[i], tmp);
207    }
208    /* Add a zero */
209    lpc2[0] = lpc[0] + QCONST16(.8f,SIG_SHIFT);
210    lpc2[1] = lpc[1] + MULT16_16_Q15(c1,lpc[0]);
211    lpc2[2] = lpc[2] + MULT16_16_Q15(c1,lpc[1]);
212    lpc2[3] = lpc[3] + MULT16_16_Q15(c1,lpc[2]);
213    lpc2[4] = MULT16_16_Q15(c1,lpc[3]);
214    celt_fir5(x_lp, lpc2, x_lp, len>>1, mem);
215 }
216 
217 /* Pure C implementation. */
218 #ifdef FIXED_POINT
219 opus_val32
220 #else
221 void
222 #endif
celt_pitch_xcorr_c(const opus_val16 * _x,const opus_val16 * _y,opus_val32 * xcorr,int len,int max_pitch,int arch)223 celt_pitch_xcorr_c(const opus_val16 *_x, const opus_val16 *_y,
224       opus_val32 *xcorr, int len, int max_pitch, int arch)
225 {
226 
227 #if 0 /* This is a simple version of the pitch correlation that should work
228          well on DSPs like Blackfin and TI C5x/C6x */
229    int i, j;
230 #ifdef FIXED_POINT
231    opus_val32 maxcorr=1;
232 #endif
233 #if !defined(OVERRIDE_PITCH_XCORR)
234    (void)arch;
235 #endif
236    for (i=0;i<max_pitch;i++)
237    {
238       opus_val32 sum = 0;
239       for (j=0;j<len;j++)
240          sum = MAC16_16(sum, _x[j], _y[i+j]);
241       xcorr[i] = sum;
242 #ifdef FIXED_POINT
243       maxcorr = MAX32(maxcorr, sum);
244 #endif
245    }
246 #ifdef FIXED_POINT
247    return maxcorr;
248 #endif
249 
250 #else /* Unrolled version of the pitch correlation -- runs faster on x86 and ARM */
251    int i;
252    /*The EDSP version requires that max_pitch is at least 1, and that _x is
253       32-bit aligned.
254      Since it's hard to put asserts in assembly, put them here.*/
255 #ifdef FIXED_POINT
256    opus_val32 maxcorr=1;
257 #endif
258    celt_assert(max_pitch>0);
259    celt_assert((((unsigned char *)_x-(unsigned char *)NULL)&3)==0);
260    for (i=0;i<max_pitch-3;i+=4)
261    {
262       opus_val32 sum[4]={0,0,0,0};
263       xcorr_kernel(_x, _y+i, sum, len, arch);
264       xcorr[i]=sum[0];
265       xcorr[i+1]=sum[1];
266       xcorr[i+2]=sum[2];
267       xcorr[i+3]=sum[3];
268 #ifdef FIXED_POINT
269       sum[0] = MAX32(sum[0], sum[1]);
270       sum[2] = MAX32(sum[2], sum[3]);
271       sum[0] = MAX32(sum[0], sum[2]);
272       maxcorr = MAX32(maxcorr, sum[0]);
273 #endif
274    }
275    /* In case max_pitch isn't a multiple of 4, do non-unrolled version. */
276    for (;i<max_pitch;i++)
277    {
278       opus_val32 sum;
279       sum = celt_inner_prod(_x, _y+i, len, arch);
280       xcorr[i] = sum;
281 #ifdef FIXED_POINT
282       maxcorr = MAX32(maxcorr, sum);
283 #endif
284    }
285 #ifdef FIXED_POINT
286    return maxcorr;
287 #endif
288 #endif
289 }
290 
pitch_search(const opus_val16 * OPUS_RESTRICT x_lp,opus_val16 * OPUS_RESTRICT y,int len,int max_pitch,int * pitch,int arch)291 void pitch_search(const opus_val16 * OPUS_RESTRICT x_lp, opus_val16 * OPUS_RESTRICT y,
292                   int len, int max_pitch, int *pitch, int arch)
293 {
294    int i, j;
295    int lag;
296    int best_pitch[2]={0,0};
297    VARDECL(opus_val16, x_lp4);
298    VARDECL(opus_val16, y_lp4);
299    VARDECL(opus_val32, xcorr);
300 #ifdef FIXED_POINT
301    opus_val32 maxcorr;
302    opus_val32 xmax, ymax;
303    int shift=0;
304 #endif
305    int offset;
306 
307    SAVE_STACK;
308 
309    celt_assert(len>0);
310    celt_assert(max_pitch>0);
311    lag = len+max_pitch;
312 
313    ALLOC(x_lp4, len>>2, opus_val16);
314    ALLOC(y_lp4, lag>>2, opus_val16);
315    ALLOC(xcorr, max_pitch>>1, opus_val32);
316 
317    /* Downsample by 2 again */
318    for (j=0;j<len>>2;j++)
319       x_lp4[j] = x_lp[2*j];
320    for (j=0;j<lag>>2;j++)
321       y_lp4[j] = y[2*j];
322 
323 #ifdef FIXED_POINT
324    xmax = celt_maxabs16(x_lp4, len>>2);
325    ymax = celt_maxabs16(y_lp4, lag>>2);
326    shift = celt_ilog2(MAX32(1, MAX32(xmax, ymax)))-11;
327    if (shift>0)
328    {
329       for (j=0;j<len>>2;j++)
330          x_lp4[j] = SHR16(x_lp4[j], shift);
331       for (j=0;j<lag>>2;j++)
332          y_lp4[j] = SHR16(y_lp4[j], shift);
333       /* Use double the shift for a MAC */
334       shift *= 2;
335    } else {
336       shift = 0;
337    }
338 #endif
339 
340    /* Coarse search with 4x decimation */
341 
342 #ifdef FIXED_POINT
343    maxcorr =
344 #endif
345    celt_pitch_xcorr(x_lp4, y_lp4, xcorr, len>>2, max_pitch>>2, arch);
346 
347    find_best_pitch(xcorr, y_lp4, len>>2, max_pitch>>2, best_pitch
348 #ifdef FIXED_POINT
349                    , 0, maxcorr
350 #endif
351                    );
352 
353    /* Finer search with 2x decimation */
354 #ifdef FIXED_POINT
355    maxcorr=1;
356 #endif
357    for (i=0;i<max_pitch>>1;i++)
358    {
359       opus_val32 sum;
360       xcorr[i] = 0;
361       if (abs(i-2*best_pitch[0])>2 && abs(i-2*best_pitch[1])>2)
362          continue;
363 #ifdef FIXED_POINT
364       sum = 0;
365       for (j=0;j<len>>1;j++)
366          sum += SHR32(MULT16_16(x_lp[j],y[i+j]), shift);
367 #else
368       sum = celt_inner_prod(x_lp, y+i, len>>1, arch);
369 #endif
370       xcorr[i] = MAX32(-1, sum);
371 #ifdef FIXED_POINT
372       maxcorr = MAX32(maxcorr, sum);
373 #endif
374    }
375    find_best_pitch(xcorr, y, len>>1, max_pitch>>1, best_pitch
376 #ifdef FIXED_POINT
377                    , shift+1, maxcorr
378 #endif
379                    );
380 
381    /* Refine by pseudo-interpolation */
382    if (best_pitch[0]>0 && best_pitch[0]<(max_pitch>>1)-1)
383    {
384       opus_val32 a, b, c;
385       a = xcorr[best_pitch[0]-1];
386       b = xcorr[best_pitch[0]];
387       c = xcorr[best_pitch[0]+1];
388       if ((c-a) > MULT16_32_Q15(QCONST16(.7f,15),b-a))
389          offset = 1;
390       else if ((a-c) > MULT16_32_Q15(QCONST16(.7f,15),b-c))
391          offset = -1;
392       else
393          offset = 0;
394    } else {
395       offset = 0;
396    }
397    *pitch = 2*best_pitch[0]-offset;
398 
399    RESTORE_STACK;
400 }
401 
402 #ifdef FIXED_POINT
compute_pitch_gain(opus_val32 xy,opus_val32 xx,opus_val32 yy)403 static opus_val16 compute_pitch_gain(opus_val32 xy, opus_val32 xx, opus_val32 yy)
404 {
405    opus_val32 x2y2;
406    int sx, sy, shift;
407    opus_val32 g;
408    opus_val16 den;
409    if (xy == 0 || xx == 0 || yy == 0)
410       return 0;
411    sx = celt_ilog2(xx)-14;
412    sy = celt_ilog2(yy)-14;
413    shift = sx + sy;
414    x2y2 = SHR32(MULT16_16(VSHR32(xx, sx), VSHR32(yy, sy)), 14);
415    if (shift & 1) {
416       if (x2y2 < 32768)
417       {
418          x2y2 <<= 1;
419          shift--;
420       } else {
421          x2y2 >>= 1;
422          shift++;
423       }
424    }
425    den = celt_rsqrt_norm(x2y2);
426    g = MULT16_32_Q15(den, xy);
427    g = VSHR32(g, (shift>>1)-1);
428    return EXTRACT16(MIN32(g, Q15ONE));
429 }
430 #else
compute_pitch_gain(opus_val32 xy,opus_val32 xx,opus_val32 yy)431 static opus_val16 compute_pitch_gain(opus_val32 xy, opus_val32 xx, opus_val32 yy)
432 {
433    return xy/celt_sqrt(1+xx*yy);
434 }
435 #endif
436 
437 static const int second_check[16] = {0, 0, 3, 2, 3, 2, 5, 2, 3, 2, 3, 2, 5, 2, 3, 2};
remove_doubling(opus_val16 * x,int maxperiod,int minperiod,int N,int * T0_,int prev_period,opus_val16 prev_gain,int arch)438 opus_val16 remove_doubling(opus_val16 *x, int maxperiod, int minperiod,
439       int N, int *T0_, int prev_period, opus_val16 prev_gain, int arch)
440 {
441    int k, i, T, T0;
442    opus_val16 g, g0;
443    opus_val16 pg;
444    opus_val32 xy,xx,yy,xy2;
445    opus_val32 xcorr[3];
446    opus_val32 best_xy, best_yy;
447    int offset;
448    int minperiod0;
449    VARDECL(opus_val32, yy_lookup);
450    SAVE_STACK;
451 
452    minperiod0 = minperiod;
453    maxperiod /= 2;
454    minperiod /= 2;
455    *T0_ /= 2;
456    prev_period /= 2;
457    N /= 2;
458    x += maxperiod;
459    if (*T0_>=maxperiod)
460       *T0_=maxperiod-1;
461 
462    T = T0 = *T0_;
463    ALLOC(yy_lookup, maxperiod+1, opus_val32);
464    dual_inner_prod(x, x, x-T0, N, &xx, &xy, arch);
465    yy_lookup[0] = xx;
466    yy=xx;
467    for (i=1;i<=maxperiod;i++)
468    {
469       yy = yy+MULT16_16(x[-i],x[-i])-MULT16_16(x[N-i],x[N-i]);
470       yy_lookup[i] = MAX32(0, yy);
471    }
472    yy = yy_lookup[T0];
473    best_xy = xy;
474    best_yy = yy;
475    g = g0 = compute_pitch_gain(xy, xx, yy);
476    /* Look for any pitch at T/k */
477    for (k=2;k<=15;k++)
478    {
479       int T1, T1b;
480       opus_val16 g1;
481       opus_val16 cont=0;
482       opus_val16 thresh;
483       T1 = celt_udiv(2*T0+k, 2*k);
484       if (T1 < minperiod)
485          break;
486       /* Look for another strong correlation at T1b */
487       if (k==2)
488       {
489          if (T1+T0>maxperiod)
490             T1b = T0;
491          else
492             T1b = T0+T1;
493       } else
494       {
495          T1b = celt_udiv(2*second_check[k]*T0+k, 2*k);
496       }
497       dual_inner_prod(x, &x[-T1], &x[-T1b], N, &xy, &xy2, arch);
498       xy = HALF32(xy + xy2);
499       yy = HALF32(yy_lookup[T1] + yy_lookup[T1b]);
500       g1 = compute_pitch_gain(xy, xx, yy);
501       if (abs(T1-prev_period)<=1)
502          cont = prev_gain;
503       else if (abs(T1-prev_period)<=2 && 5*k*k < T0)
504          cont = HALF16(prev_gain);
505       else
506          cont = 0;
507       thresh = MAX16(QCONST16(.3f,15), MULT16_16_Q15(QCONST16(.7f,15),g0)-cont);
508       /* Bias against very high pitch (very short period) to avoid false-positives
509          due to short-term correlation */
510       if (T1<3*minperiod)
511          thresh = MAX16(QCONST16(.4f,15), MULT16_16_Q15(QCONST16(.85f,15),g0)-cont);
512       else if (T1<2*minperiod)
513          thresh = MAX16(QCONST16(.5f,15), MULT16_16_Q15(QCONST16(.9f,15),g0)-cont);
514       if (g1 > thresh)
515       {
516          best_xy = xy;
517          best_yy = yy;
518          T = T1;
519          g = g1;
520       }
521    }
522    best_xy = MAX32(0, best_xy);
523    if (best_yy <= best_xy)
524       pg = Q15ONE;
525    else
526       pg = SHR32(frac_div32(best_xy,best_yy+1),16);
527 
528    for (k=0;k<3;k++)
529       xcorr[k] = celt_inner_prod(x, x-(T+k-1), N, arch);
530    if ((xcorr[2]-xcorr[0]) > MULT16_32_Q15(QCONST16(.7f,15),xcorr[1]-xcorr[0]))
531       offset = 1;
532    else if ((xcorr[0]-xcorr[2]) > MULT16_32_Q15(QCONST16(.7f,15),xcorr[1]-xcorr[2]))
533       offset = -1;
534    else
535       offset = 0;
536    if (pg > g)
537       pg = g;
538    *T0_ = 2*T+offset;
539 
540    if (*T0_<minperiod0)
541       *T0_=minperiod0;
542    RESTORE_STACK;
543    return pg;
544 }
545