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