• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  *  Copyright (c) 2011 The WebRTC project authors. All Rights Reserved.
3  *
4  *  Use of this source code is governed by a BSD-style license
5  *  that can be found in the LICENSE file in the root of the source
6  *  tree. An additional intellectual property rights grant can be found
7  *  in the file PATENTS.  All contributing project authors may
8  *  be found in the AUTHORS file in the root of the source tree.
9  */
10 
11 #include "modules/audio_coding/codecs/isac/fix/source/pitch_estimator.h"
12 #include "common_audio/signal_processing/include/signal_processing_library.h"
13 #include "rtc_base/compile_assert_c.h"
14 
15 /* log2[0.2, 0.5, 0.98] in Q8 */
16 static const int16_t kLogLagWinQ8[3] = {
17   -594, -256, -7
18 };
19 
20 /* [1 -0.75 0.25] in Q12 */
21 static const int16_t kACoefQ12[3] = {
22   4096, -3072, 1024
23 };
24 
WebRtcIsacfix_Log2Q8(uint32_t x)25 int32_t WebRtcIsacfix_Log2Q8(uint32_t x) {
26   int32_t zeros;
27   int16_t frac;
28 
29   zeros=WebRtcSpl_NormU32(x);
30   frac = (int16_t)(((x << zeros) & 0x7FFFFFFF) >> 23);
31   /* log2(magn(i)) */
32 
33   return ((31 - zeros) << 8) + frac;
34 }
35 
Exp2Q10(int16_t x)36 static __inline int16_t Exp2Q10(int16_t x) { // Both in and out in Q10
37 
38   int16_t tmp16_1, tmp16_2;
39 
40   tmp16_2=(int16_t)(0x0400|(x&0x03FF));
41   tmp16_1 = -(x >> 10);
42   if(tmp16_1>0)
43     return tmp16_2 >> tmp16_1;
44   else
45     return tmp16_2 << -tmp16_1;
46 
47 }
48 
49 
50 
51 /* 1D parabolic interpolation . All input and output values are in Q8 */
Intrp1DQ8(int32_t * x,int32_t * fx,int32_t * y,int32_t * fy)52 static __inline void Intrp1DQ8(int32_t *x, int32_t *fx, int32_t *y, int32_t *fy) {
53 
54   int16_t sign1=1, sign2=1;
55   int32_t r32, q32, t32, nom32, den32;
56   int16_t t16, tmp16, tmp16_1;
57 
58   if ((fx[0]>0) && (fx[2]>0)) {
59     r32=fx[1]-fx[2];
60     q32=fx[0]-fx[1];
61     nom32=q32+r32;
62     den32 = (q32 - r32) * 2;
63     if (nom32<0)
64       sign1=-1;
65     if (den32<0)
66       sign2=-1;
67 
68     /* t = (q32+r32)/(2*(q32-r32)) = (fx[0]-fx[1] + fx[1]-fx[2])/(2 * fx[0]-fx[1] - (fx[1]-fx[2]))*/
69     /* (Signs are removed because WebRtcSpl_DivResultInQ31 can't handle negative numbers) */
70     /* t in Q31, without signs */
71     t32 = WebRtcSpl_DivResultInQ31(nom32 * sign1, den32 * sign2);
72 
73     t16 = (int16_t)(t32 >> 23);  /* Q8 */
74     t16=t16*sign1*sign2;        /* t in Q8 with signs */
75 
76     *y = x[0]+t16;          /* Q8 */
77     // *y = x[1]+t16;          /* Q8 */
78 
79     /* The following code calculates fy in three steps */
80     /* fy = 0.5 * t * (t-1) * fx[0] + (1-t*t) * fx[1] + 0.5 * t * (t+1) * fx[2]; */
81 
82     /* Part I: 0.5 * t * (t-1) * fx[0] */
83     tmp16_1 = (int16_t)(t16 * t16);  /* Q8*Q8=Q16 */
84     tmp16_1 >>= 2;  /* Q16>>2 = Q14 */
85     t16 <<= 6;  /* Q8<<6 = Q14  */
86     tmp16 = tmp16_1-t16;
87     *fy = WEBRTC_SPL_MUL_16_32_RSFT15(tmp16, fx[0]); /* (Q14 * Q8 >>15)/2 = Q8 */
88 
89     /* Part II: (1-t*t) * fx[1] */
90     tmp16 = 16384-tmp16_1;        /* 1 in Q14 - Q14 */
91     *fy += WEBRTC_SPL_MUL_16_32_RSFT14(tmp16, fx[1]);/* Q14 * Q8 >> 14 = Q8 */
92 
93     /* Part III: 0.5 * t * (t+1) * fx[2] */
94     tmp16 = tmp16_1+t16;
95     *fy += WEBRTC_SPL_MUL_16_32_RSFT15(tmp16, fx[2]);/* (Q14 * Q8 >>15)/2 = Q8 */
96   } else {
97     *y = x[0];
98     *fy= fx[1];
99   }
100 }
101 
102 
FindFour32(int32_t * in,int16_t length,int16_t * bestind)103 static void FindFour32(int32_t *in, int16_t length, int16_t *bestind)
104 {
105   int32_t best[4]= {-100, -100, -100, -100};
106   int16_t k;
107 
108   for (k=0; k<length; k++) {
109     if (in[k] > best[3]) {
110       if (in[k] > best[2]) {
111         if (in[k] > best[1]) {
112           if (in[k] > best[0]) { // The Best
113             best[3] = best[2];
114             bestind[3] = bestind[2];
115             best[2] = best[1];
116             bestind[2] = bestind[1];
117             best[1] = best[0];
118             bestind[1] = bestind[0];
119             best[0] = in[k];
120             bestind[0] = k;
121           } else { // 2nd best
122             best[3] = best[2];
123             bestind[3] = bestind[2];
124             best[2] = best[1];
125             bestind[2] = bestind[1];
126             best[1] = in[k];
127             bestind[1] = k;
128           }
129         } else { // 3rd best
130           best[3] = best[2];
131           bestind[3] = bestind[2];
132           best[2] = in[k];
133           bestind[2] = k;
134         }
135       } else {  // 4th best
136         best[3] = in[k];
137         bestind[3] = k;
138       }
139     }
140   }
141 }
142 
143 
144 
145 
146 
147 extern void WebRtcIsacfix_PCorr2Q32(const int16_t *in, int32_t *logcorQ8);
148 
149 
150 
WebRtcIsacfix_InitialPitch(const int16_t * in,PitchAnalysisStruct * State,int16_t * lagsQ7)151 void WebRtcIsacfix_InitialPitch(const int16_t *in, /* Q0 */
152                                 PitchAnalysisStruct *State,
153                                 int16_t *lagsQ7                   /* Q7 */
154                                 )
155 {
156   int16_t buf_dec16[PITCH_CORR_LEN2+PITCH_CORR_STEP2+PITCH_MAX_LAG/2+2];
157   int32_t *crrvecQ8_1,*crrvecQ8_2;
158   int32_t cv1q[PITCH_LAG_SPAN2+2],cv2q[PITCH_LAG_SPAN2+2], peakvq[PITCH_LAG_SPAN2+2];
159   int k;
160   int16_t peaks_indq;
161   int16_t peakiq[PITCH_LAG_SPAN2];
162   int32_t corr;
163   int32_t corr32, corr_max32, corr_max_o32;
164   int16_t npkq;
165   int16_t best4q[4]={0,0,0,0};
166   int32_t xq[3],yq[1],fyq[1];
167   int32_t *fxq;
168   int32_t best_lag1q, best_lag2q;
169   int32_t tmp32a,tmp32b,lag32,ratq;
170   int16_t start;
171   int16_t oldgQ12, tmp16a, tmp16b, gain_bias16,tmp16c, tmp16d, bias16;
172   int32_t tmp32c,tmp32d, tmp32e;
173   int16_t old_lagQ;
174   int32_t old_lagQ8;
175   int32_t lagsQ8[4];
176 
177   old_lagQ = State->PFstr_wght.oldlagQ7; // Q7
178   old_lagQ8 = old_lagQ << 1;  // Q8
179 
180   oldgQ12= State->PFstr_wght.oldgainQ12;
181 
182   crrvecQ8_1=&cv1q[1];
183   crrvecQ8_2=&cv2q[1];
184 
185 
186   /* copy old values from state buffer */
187   memcpy(buf_dec16, State->dec_buffer16, sizeof(State->dec_buffer16));
188 
189   /* decimation; put result after the old values */
190   WebRtcIsacfix_DecimateAllpass32(in, State->decimator_state32, PITCH_FRAME_LEN,
191                                   &buf_dec16[PITCH_CORR_LEN2+PITCH_CORR_STEP2+PITCH_MAX_LAG/2-PITCH_FRAME_LEN/2+2]);
192 
193   /* low-pass filtering */
194   start= PITCH_CORR_LEN2+PITCH_CORR_STEP2+PITCH_MAX_LAG/2-PITCH_FRAME_LEN/2+2;
195   WebRtcSpl_FilterARFastQ12(&buf_dec16[start],&buf_dec16[start],(int16_t*)kACoefQ12,3, PITCH_FRAME_LEN/2);
196 
197   /* copy end part back into state buffer */
198   for (k = 0; k < (PITCH_CORR_LEN2+PITCH_CORR_STEP2+PITCH_MAX_LAG/2-PITCH_FRAME_LEN/2+2); k++)
199     State->dec_buffer16[k] = buf_dec16[k+PITCH_FRAME_LEN/2];
200 
201 
202   /* compute correlation for first and second half of the frame */
203   WebRtcIsacfix_PCorr2Q32(buf_dec16, crrvecQ8_1);
204   WebRtcIsacfix_PCorr2Q32(buf_dec16 + PITCH_CORR_STEP2, crrvecQ8_2);
205 
206 
207   /* bias towards pitch lag of previous frame */
208   tmp32a = WebRtcIsacfix_Log2Q8((uint32_t) old_lagQ8) - 2304;
209       // log2(0.5*oldlag) in Q8
210   tmp32b = oldgQ12 * oldgQ12 >> 10;  // Q12 & * 4.0;
211   gain_bias16 = (int16_t) tmp32b;  //Q12
212   if (gain_bias16 > 3276) gain_bias16 = 3276; // 0.8 in Q12
213 
214 
215   for (k = 0; k < PITCH_LAG_SPAN2; k++)
216   {
217     if (crrvecQ8_1[k]>0) {
218       tmp32b = WebRtcIsacfix_Log2Q8((uint32_t) (k + (PITCH_MIN_LAG/2-2)));
219       tmp16a = (int16_t) (tmp32b - tmp32a); // Q8 & fabs(ratio)<4
220       tmp32c = tmp16a * tmp16a >> 6;  // Q10
221       tmp16b = (int16_t) tmp32c; // Q10 & <8
222       tmp32d = tmp16b * 177 >> 8;  // mult with ln2 in Q8
223       tmp16c = (int16_t) tmp32d; // Q10 & <4
224       tmp16d = Exp2Q10((int16_t) -tmp16c); //Q10
225       tmp32c = gain_bias16 * tmp16d >> 13;  // Q10  & * 0.5
226       bias16 = (int16_t) (1024 + tmp32c); // Q10
227       tmp32b = WebRtcIsacfix_Log2Q8((uint32_t)bias16) - 2560;
228           // Q10 in -> Q8 out with 10*2^8 offset
229       crrvecQ8_1[k] += tmp32b ; // -10*2^8 offset
230     }
231   }
232 
233   /* taper correlation functions */
234   for (k = 0; k < 3; k++) {
235     crrvecQ8_1[k] += kLogLagWinQ8[k];
236     crrvecQ8_2[k] += kLogLagWinQ8[k];
237 
238     crrvecQ8_1[PITCH_LAG_SPAN2-1-k] += kLogLagWinQ8[k];
239     crrvecQ8_2[PITCH_LAG_SPAN2-1-k] += kLogLagWinQ8[k];
240   }
241 
242 
243   /* Make zeropadded corr vectors */
244   cv1q[0]=0;
245   cv2q[0]=0;
246   cv1q[PITCH_LAG_SPAN2+1]=0;
247   cv2q[PITCH_LAG_SPAN2+1]=0;
248   corr_max32 = 0;
249 
250   for (k = 1; k <= PITCH_LAG_SPAN2; k++)
251   {
252 
253 
254     corr32=crrvecQ8_1[k-1];
255     if (corr32 > corr_max32)
256       corr_max32 = corr32;
257 
258     corr32=crrvecQ8_2[k-1];
259     corr32 += -4; // Compensate for later (log2(0.99))
260 
261     if (corr32 > corr_max32)
262       corr_max32 = corr32;
263 
264   }
265 
266   /* threshold value to qualify as a peak */
267   // corr_max32 += -726; // log(0.14)/log(2.0) in Q8
268   corr_max32 += -1000; // log(0.14)/log(2.0) in Q8
269   corr_max_o32 = corr_max32;
270 
271 
272   /* find peaks in corr1 */
273   peaks_indq = 0;
274   for (k = 1; k <= PITCH_LAG_SPAN2; k++)
275   {
276     corr32=cv1q[k];
277     if (corr32>corr_max32) { // Disregard small peaks
278       if ((corr32>=cv1q[k-1]) && (corr32>cv1q[k+1])) { // Peak?
279         peakvq[peaks_indq] = corr32;
280         peakiq[peaks_indq++] = k;
281       }
282     }
283   }
284 
285 
286   /* find highest interpolated peak */
287   corr_max32=0;
288   best_lag1q =0;
289   if (peaks_indq > 0) {
290     FindFour32(peakvq, (int16_t) peaks_indq, best4q);
291     npkq = WEBRTC_SPL_MIN(peaks_indq, 4);
292 
293     for (k=0;k<npkq;k++) {
294 
295       lag32 =  peakiq[best4q[k]];
296       fxq = &cv1q[peakiq[best4q[k]]-1];
297       xq[0]= lag32;
298       xq[0] <<= 8;
299       Intrp1DQ8(xq, fxq, yq, fyq);
300 
301       tmp32a= WebRtcIsacfix_Log2Q8((uint32_t) *yq) - 2048; // offset 8*2^8
302       /* Bias towards short lags */
303       /* log(pow(0.8, log(2.0 * *y )))/log(2.0) */
304       tmp32b = (int16_t)tmp32a * -42 >> 8;
305       tmp32c= tmp32b + 256;
306       *fyq += tmp32c;
307       if (*fyq > corr_max32) {
308         corr_max32 = *fyq;
309         best_lag1q = *yq;
310       }
311     }
312     tmp32b = (best_lag1q - OFFSET_Q8) * 2;
313     lagsQ8[0] = tmp32b + PITCH_MIN_LAG_Q8;
314     lagsQ8[1] = lagsQ8[0];
315   } else {
316     lagsQ8[0] = old_lagQ8;
317     lagsQ8[1] = lagsQ8[0];
318   }
319 
320   /* Bias towards constant pitch */
321   tmp32a = lagsQ8[0] - PITCH_MIN_LAG_Q8;
322   ratq = (tmp32a >> 1) + OFFSET_Q8;
323 
324   for (k = 1; k <= PITCH_LAG_SPAN2; k++)
325   {
326     tmp32a = k << 7; // 0.5*k Q8
327     tmp32b = tmp32a * 2 - ratq;  // Q8
328     tmp32c = (int16_t)tmp32b * (int16_t)tmp32b >> 8;  // Q8
329 
330     tmp32b = tmp32c + (ratq >> 1);
331         // (k-r)^2 + 0.5 * r  Q8
332     tmp32c = WebRtcIsacfix_Log2Q8((uint32_t)tmp32a) - 2048;
333         // offset 8*2^8 , log2(0.5*k) Q8
334     tmp32d = WebRtcIsacfix_Log2Q8((uint32_t)tmp32b) - 2048;
335         // offset 8*2^8 , log2(0.5*k) Q8
336     tmp32e =  tmp32c - tmp32d;
337 
338     cv2q[k] += tmp32e >> 1;
339 
340   }
341 
342   /* find peaks in corr2 */
343   corr_max32 = corr_max_o32;
344   peaks_indq = 0;
345 
346   for (k = 1; k <= PITCH_LAG_SPAN2; k++)
347   {
348     corr=cv2q[k];
349     if (corr>corr_max32) { // Disregard small peaks
350       if ((corr>=cv2q[k-1]) && (corr>cv2q[k+1])) { // Peak?
351         peakvq[peaks_indq] = corr;
352         peakiq[peaks_indq++] = k;
353       }
354     }
355   }
356 
357 
358 
359   /* find highest interpolated peak */
360   corr_max32 = 0;
361   best_lag2q =0;
362   if (peaks_indq > 0) {
363 
364     FindFour32(peakvq, (int16_t) peaks_indq, best4q);
365     npkq = WEBRTC_SPL_MIN(peaks_indq, 4);
366     for (k=0;k<npkq;k++) {
367 
368       lag32 =  peakiq[best4q[k]];
369       fxq = &cv2q[peakiq[best4q[k]]-1];
370 
371       xq[0]= lag32;
372       xq[0] <<= 8;
373       Intrp1DQ8(xq, fxq, yq, fyq);
374 
375       /* Bias towards short lags */
376       /* log(pow(0.8, log(2.0f * *y )))/log(2.0f) */
377       tmp32a= WebRtcIsacfix_Log2Q8((uint32_t) *yq) - 2048; // offset 8*2^8
378       tmp32b = (int16_t)tmp32a * -82 >> 8;
379       tmp32c= tmp32b + 256;
380       *fyq += tmp32c;
381       if (*fyq > corr_max32) {
382         corr_max32 = *fyq;
383         best_lag2q = *yq;
384       }
385     }
386 
387     tmp32b = (best_lag2q - OFFSET_Q8) * 2;
388     lagsQ8[2] = tmp32b + PITCH_MIN_LAG_Q8;
389     lagsQ8[3] = lagsQ8[2];
390   } else {
391     lagsQ8[2] = lagsQ8[0];
392     lagsQ8[3] = lagsQ8[0];
393   }
394 
395   lagsQ7[0] = (int16_t)(lagsQ8[0] >> 1);
396   lagsQ7[1] = (int16_t)(lagsQ8[1] >> 1);
397   lagsQ7[2] = (int16_t)(lagsQ8[2] >> 1);
398   lagsQ7[3] = (int16_t)(lagsQ8[3] >> 1);
399 }
400 
401 
402 
WebRtcIsacfix_PitchAnalysis(const int16_t * inn,int16_t * outQ0,PitchAnalysisStruct * State,int16_t * PitchLags_Q7,int16_t * PitchGains_Q12)403 void WebRtcIsacfix_PitchAnalysis(const int16_t *inn,               /* PITCH_FRAME_LEN samples */
404                                  int16_t *outQ0,                  /* PITCH_FRAME_LEN+QLOOKAHEAD samples */
405                                  PitchAnalysisStruct *State,
406                                  int16_t *PitchLags_Q7,
407                                  int16_t *PitchGains_Q12)
408 {
409   int16_t inbufQ0[PITCH_FRAME_LEN + QLOOKAHEAD];
410   int16_t k;
411 
412   /* inital pitch estimate */
413   WebRtcIsacfix_InitialPitch(inn, State,  PitchLags_Q7);
414 
415 
416   /* Calculate gain */
417   WebRtcIsacfix_PitchFilterGains(inn, &(State->PFstr_wght), PitchLags_Q7, PitchGains_Q12);
418 
419   /* concatenate previous input's end and current input */
420   for (k = 0; k < QLOOKAHEAD; k++) {
421     inbufQ0[k] = State->inbuf[k];
422   }
423   for (k = 0; k < PITCH_FRAME_LEN; k++) {
424     inbufQ0[k+QLOOKAHEAD] = (int16_t) inn[k];
425   }
426 
427   /* lookahead pitch filtering for masking analysis */
428   WebRtcIsacfix_PitchFilter(inbufQ0, outQ0, &(State->PFstr), PitchLags_Q7,PitchGains_Q12, 2);
429 
430 
431   /* store last part of input */
432   for (k = 0; k < QLOOKAHEAD; k++) {
433     State->inbuf[k] = inbufQ0[k + PITCH_FRAME_LEN];
434   }
435 }
436