• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * http://www.kurims.kyoto-u.ac.jp/~ooura/fft.html
3  * Copyright Takuya OOURA, 1996-2001
4  *
5  * You may use, copy, modify and distribute this code for any purpose (include
6  * commercial use) and without fee. Please refer to this package when you modify
7  * this code.
8  *
9  * Changes:
10  * Trivial type modifications by the WebRTC authors.
11  */
12 
13 /*
14 Fast Fourier/Cosine/Sine Transform
15     dimension   :one
16     data length :power of 2
17     decimation  :frequency
18     radix       :4, 2
19     data        :inplace
20     table       :use
21 functions
22     cdft: Complex Discrete Fourier Transform
23     rdft: Real Discrete Fourier Transform
24     ddct: Discrete Cosine Transform
25     ddst: Discrete Sine Transform
26     dfct: Cosine Transform of RDFT (Real Symmetric DFT)
27     dfst: Sine Transform of RDFT (Real Anti-symmetric DFT)
28 function prototypes
29     void cdft(int, int, float *, int *, float *);
30     void rdft(size_t, int, float *, size_t *, float *);
31     void ddct(int, int, float *, int *, float *);
32     void ddst(int, int, float *, int *, float *);
33     void dfct(int, float *, float *, int *, float *);
34     void dfst(int, float *, float *, int *, float *);
35 
36 
37 -------- Complex DFT (Discrete Fourier Transform) --------
38     [definition]
39         <case1>
40             X[k] = sum_j=0^n-1 x[j]*exp(2*pi*i*j*k/n), 0<=k<n
41         <case2>
42             X[k] = sum_j=0^n-1 x[j]*exp(-2*pi*i*j*k/n), 0<=k<n
43         (notes: sum_j=0^n-1 is a summation from j=0 to n-1)
44     [usage]
45         <case1>
46             ip[0] = 0; // first time only
47             cdft(2*n, 1, a, ip, w);
48         <case2>
49             ip[0] = 0; // first time only
50             cdft(2*n, -1, a, ip, w);
51     [parameters]
52         2*n            :data length (int)
53                         n >= 1, n = power of 2
54         a[0...2*n-1]   :input/output data (float *)
55                         input data
56                             a[2*j] = Re(x[j]),
57                             a[2*j+1] = Im(x[j]), 0<=j<n
58                         output data
59                             a[2*k] = Re(X[k]),
60                             a[2*k+1] = Im(X[k]), 0<=k<n
61         ip[0...*]      :work area for bit reversal (int *)
62                         length of ip >= 2+sqrt(n)
63                         strictly,
64                         length of ip >=
65                             2+(1<<(int)(log(n+0.5)/log(2))/2).
66                         ip[0],ip[1] are pointers of the cos/sin table.
67         w[0...n/2-1]   :cos/sin table (float *)
68                         w[],ip[] are initialized if ip[0] == 0.
69     [remark]
70         Inverse of
71             cdft(2*n, -1, a, ip, w);
72         is
73             cdft(2*n, 1, a, ip, w);
74             for (j = 0; j <= 2 * n - 1; j++) {
75                 a[j] *= 1.0 / n;
76             }
77         .
78 
79 
80 -------- Real DFT / Inverse of Real DFT --------
81     [definition]
82         <case1> RDFT
83             R[k] = sum_j=0^n-1 a[j]*cos(2*pi*j*k/n), 0<=k<=n/2
84             I[k] = sum_j=0^n-1 a[j]*sin(2*pi*j*k/n), 0<k<n/2
85         <case2> IRDFT (excluding scale)
86             a[k] = (R[0] + R[n/2]*cos(pi*k))/2 +
87                    sum_j=1^n/2-1 R[j]*cos(2*pi*j*k/n) +
88                    sum_j=1^n/2-1 I[j]*sin(2*pi*j*k/n), 0<=k<n
89     [usage]
90         <case1>
91             ip[0] = 0; // first time only
92             rdft(n, 1, a, ip, w);
93         <case2>
94             ip[0] = 0; // first time only
95             rdft(n, -1, a, ip, w);
96     [parameters]
97         n              :data length (size_t)
98                         n >= 2, n = power of 2
99         a[0...n-1]     :input/output data (float *)
100                         <case1>
101                             output data
102                                 a[2*k] = R[k], 0<=k<n/2
103                                 a[2*k+1] = I[k], 0<k<n/2
104                                 a[1] = R[n/2]
105                         <case2>
106                             input data
107                                 a[2*j] = R[j], 0<=j<n/2
108                                 a[2*j+1] = I[j], 0<j<n/2
109                                 a[1] = R[n/2]
110         ip[0...*]      :work area for bit reversal (size_t *)
111                         length of ip >= 2+sqrt(n/2)
112                         strictly,
113                         length of ip >=
114                             2+(1<<(int)(log(n/2+0.5)/log(2))/2).
115                         ip[0],ip[1] are pointers of the cos/sin table.
116         w[0...n/2-1]   :cos/sin table (float *)
117                         w[],ip[] are initialized if ip[0] == 0.
118     [remark]
119         Inverse of
120             rdft(n, 1, a, ip, w);
121         is
122             rdft(n, -1, a, ip, w);
123             for (j = 0; j <= n - 1; j++) {
124                 a[j] *= 2.0 / n;
125             }
126         .
127 
128 
129 -------- DCT (Discrete Cosine Transform) / Inverse of DCT --------
130     [definition]
131         <case1> IDCT (excluding scale)
132             C[k] = sum_j=0^n-1 a[j]*cos(pi*j*(k+1/2)/n), 0<=k<n
133         <case2> DCT
134             C[k] = sum_j=0^n-1 a[j]*cos(pi*(j+1/2)*k/n), 0<=k<n
135     [usage]
136         <case1>
137             ip[0] = 0; // first time only
138             ddct(n, 1, a, ip, w);
139         <case2>
140             ip[0] = 0; // first time only
141             ddct(n, -1, a, ip, w);
142     [parameters]
143         n              :data length (int)
144                         n >= 2, n = power of 2
145         a[0...n-1]     :input/output data (float *)
146                         output data
147                             a[k] = C[k], 0<=k<n
148         ip[0...*]      :work area for bit reversal (int *)
149                         length of ip >= 2+sqrt(n/2)
150                         strictly,
151                         length of ip >=
152                             2+(1<<(int)(log(n/2+0.5)/log(2))/2).
153                         ip[0],ip[1] are pointers of the cos/sin table.
154         w[0...n*5/4-1] :cos/sin table (float *)
155                         w[],ip[] are initialized if ip[0] == 0.
156     [remark]
157         Inverse of
158             ddct(n, -1, a, ip, w);
159         is
160             a[0] *= 0.5;
161             ddct(n, 1, a, ip, w);
162             for (j = 0; j <= n - 1; j++) {
163                 a[j] *= 2.0 / n;
164             }
165         .
166 
167 
168 -------- DST (Discrete Sine Transform) / Inverse of DST --------
169     [definition]
170         <case1> IDST (excluding scale)
171             S[k] = sum_j=1^n A[j]*sin(pi*j*(k+1/2)/n), 0<=k<n
172         <case2> DST
173             S[k] = sum_j=0^n-1 a[j]*sin(pi*(j+1/2)*k/n), 0<k<=n
174     [usage]
175         <case1>
176             ip[0] = 0; // first time only
177             ddst(n, 1, a, ip, w);
178         <case2>
179             ip[0] = 0; // first time only
180             ddst(n, -1, a, ip, w);
181     [parameters]
182         n              :data length (int)
183                         n >= 2, n = power of 2
184         a[0...n-1]     :input/output data (float *)
185                         <case1>
186                             input data
187                                 a[j] = A[j], 0<j<n
188                                 a[0] = A[n]
189                             output data
190                                 a[k] = S[k], 0<=k<n
191                         <case2>
192                             output data
193                                 a[k] = S[k], 0<k<n
194                                 a[0] = S[n]
195         ip[0...*]      :work area for bit reversal (int *)
196                         length of ip >= 2+sqrt(n/2)
197                         strictly,
198                         length of ip >=
199                             2+(1<<(int)(log(n/2+0.5)/log(2))/2).
200                         ip[0],ip[1] are pointers of the cos/sin table.
201         w[0...n*5/4-1] :cos/sin table (float *)
202                         w[],ip[] are initialized if ip[0] == 0.
203     [remark]
204         Inverse of
205             ddst(n, -1, a, ip, w);
206         is
207             a[0] *= 0.5;
208             ddst(n, 1, a, ip, w);
209             for (j = 0; j <= n - 1; j++) {
210                 a[j] *= 2.0 / n;
211             }
212         .
213 
214 
215 -------- Cosine Transform of RDFT (Real Symmetric DFT) --------
216     [definition]
217         C[k] = sum_j=0^n a[j]*cos(pi*j*k/n), 0<=k<=n
218     [usage]
219         ip[0] = 0; // first time only
220         dfct(n, a, t, ip, w);
221     [parameters]
222         n              :data length - 1 (int)
223                         n >= 2, n = power of 2
224         a[0...n]       :input/output data (float *)
225                         output data
226                             a[k] = C[k], 0<=k<=n
227         t[0...n/2]     :work area (float *)
228         ip[0...*]      :work area for bit reversal (int *)
229                         length of ip >= 2+sqrt(n/4)
230                         strictly,
231                         length of ip >=
232                             2+(1<<(int)(log(n/4+0.5)/log(2))/2).
233                         ip[0],ip[1] are pointers of the cos/sin table.
234         w[0...n*5/8-1] :cos/sin table (float *)
235                         w[],ip[] are initialized if ip[0] == 0.
236     [remark]
237         Inverse of
238             a[0] *= 0.5;
239             a[n] *= 0.5;
240             dfct(n, a, t, ip, w);
241         is
242             a[0] *= 0.5;
243             a[n] *= 0.5;
244             dfct(n, a, t, ip, w);
245             for (j = 0; j <= n; j++) {
246                 a[j] *= 2.0 / n;
247             }
248         .
249 
250 
251 -------- Sine Transform of RDFT (Real Anti-symmetric DFT) --------
252     [definition]
253         S[k] = sum_j=1^n-1 a[j]*sin(pi*j*k/n), 0<k<n
254     [usage]
255         ip[0] = 0; // first time only
256         dfst(n, a, t, ip, w);
257     [parameters]
258         n              :data length + 1 (int)
259                         n >= 2, n = power of 2
260         a[0...n-1]     :input/output data (float *)
261                         output data
262                             a[k] = S[k], 0<k<n
263                         (a[0] is used for work area)
264         t[0...n/2-1]   :work area (float *)
265         ip[0...*]      :work area for bit reversal (int *)
266                         length of ip >= 2+sqrt(n/4)
267                         strictly,
268                         length of ip >=
269                             2+(1<<(int)(log(n/4+0.5)/log(2))/2).
270                         ip[0],ip[1] are pointers of the cos/sin table.
271         w[0...n*5/8-1] :cos/sin table (float *)
272                         w[],ip[] are initialized if ip[0] == 0.
273     [remark]
274         Inverse of
275             dfst(n, a, t, ip, w);
276         is
277             dfst(n, a, t, ip, w);
278             for (j = 1; j <= n - 1; j++) {
279                 a[j] *= 2.0 / n;
280             }
281         .
282 
283 
284 Appendix :
285     The cos/sin table is recalculated when the larger table required.
286     w[] and ip[] are compatible with all routines.
287 */
288 
289 #include <math.h>
290 #include <stddef.h>
291 
292 #include "common_audio/third_party/ooura/fft_size_256/fft4g.h"
293 
294 namespace webrtc {
295 
296 namespace {
297 
298 void makewt(size_t nw, size_t* ip, float* w);
299 void makect(size_t nc, size_t* ip, float* c);
300 void bitrv2(size_t n, size_t* ip, float* a);
301 void cftfsub(size_t n, float* a, float* w);
302 void cftbsub(size_t n, float* a, float* w);
303 void cft1st(size_t n, float* a, float* w);
304 void cftmdl(size_t n, size_t l, float* a, float* w);
305 void rftfsub(size_t n, float* a, size_t nc, float* c);
306 void rftbsub(size_t n, float* a, size_t nc, float* c);
307 
308 /* -------- initializing routines -------- */
309 
makewt(size_t nw,size_t * ip,float * w)310 void makewt(size_t nw, size_t* ip, float* w) {
311   size_t j, nwh;
312   float delta, x, y;
313 
314   ip[0] = nw;
315   ip[1] = 1;
316   if (nw > 2) {
317     nwh = nw >> 1;
318     delta = atanf(1.0f) / nwh;
319     w[0] = 1;
320     w[1] = 0;
321     w[nwh] = (float)cos(delta * nwh);
322     w[nwh + 1] = w[nwh];
323     if (nwh > 2) {
324       for (j = 2; j < nwh; j += 2) {
325         x = (float)cos(delta * j);
326         y = (float)sin(delta * j);
327         w[j] = x;
328         w[j + 1] = y;
329         w[nw - j] = y;
330         w[nw - j + 1] = x;
331       }
332       bitrv2(nw, ip + 2, w);
333     }
334   }
335 }
336 
makect(size_t nc,size_t * ip,float * c)337 void makect(size_t nc, size_t* ip, float* c) {
338   size_t j, nch;
339   float delta;
340 
341   ip[1] = nc;
342   if (nc > 1) {
343     nch = nc >> 1;
344     delta = atanf(1.0f) / nch;
345     c[0] = (float)cos(delta * nch);
346     c[nch] = 0.5f * c[0];
347     for (j = 1; j < nch; j++) {
348       c[j] = 0.5f * (float)cos(delta * j);
349       c[nc - j] = 0.5f * (float)sin(delta * j);
350     }
351   }
352 }
353 
354 /* -------- child routines -------- */
355 
bitrv2(size_t n,size_t * ip,float * a)356 void bitrv2(size_t n, size_t* ip, float* a) {
357   size_t j, j1, k, k1, l, m, m2;
358   float xr, xi, yr, yi;
359 
360   ip[0] = 0;
361   l = n;
362   m = 1;
363   while ((m << 3) < l) {
364     l >>= 1;
365     for (j = 0; j < m; j++) {
366       ip[m + j] = ip[j] + l;
367     }
368     m <<= 1;
369   }
370   m2 = 2 * m;
371   if ((m << 3) == l) {
372     for (k = 0; k < m; k++) {
373       for (j = 0; j < k; j++) {
374         j1 = 2 * j + ip[k];
375         k1 = 2 * k + ip[j];
376         xr = a[j1];
377         xi = a[j1 + 1];
378         yr = a[k1];
379         yi = a[k1 + 1];
380         a[j1] = yr;
381         a[j1 + 1] = yi;
382         a[k1] = xr;
383         a[k1 + 1] = xi;
384         j1 += m2;
385         k1 += 2 * m2;
386         xr = a[j1];
387         xi = a[j1 + 1];
388         yr = a[k1];
389         yi = a[k1 + 1];
390         a[j1] = yr;
391         a[j1 + 1] = yi;
392         a[k1] = xr;
393         a[k1 + 1] = xi;
394         j1 += m2;
395         k1 -= m2;
396         xr = a[j1];
397         xi = a[j1 + 1];
398         yr = a[k1];
399         yi = a[k1 + 1];
400         a[j1] = yr;
401         a[j1 + 1] = yi;
402         a[k1] = xr;
403         a[k1 + 1] = xi;
404         j1 += m2;
405         k1 += 2 * m2;
406         xr = a[j1];
407         xi = a[j1 + 1];
408         yr = a[k1];
409         yi = a[k1 + 1];
410         a[j1] = yr;
411         a[j1 + 1] = yi;
412         a[k1] = xr;
413         a[k1 + 1] = xi;
414       }
415       j1 = 2 * k + m2 + ip[k];
416       k1 = j1 + m2;
417       xr = a[j1];
418       xi = a[j1 + 1];
419       yr = a[k1];
420       yi = a[k1 + 1];
421       a[j1] = yr;
422       a[j1 + 1] = yi;
423       a[k1] = xr;
424       a[k1 + 1] = xi;
425     }
426   } else {
427     for (k = 1; k < m; k++) {
428       for (j = 0; j < k; j++) {
429         j1 = 2 * j + ip[k];
430         k1 = 2 * k + ip[j];
431         xr = a[j1];
432         xi = a[j1 + 1];
433         yr = a[k1];
434         yi = a[k1 + 1];
435         a[j1] = yr;
436         a[j1 + 1] = yi;
437         a[k1] = xr;
438         a[k1 + 1] = xi;
439         j1 += m2;
440         k1 += m2;
441         xr = a[j1];
442         xi = a[j1 + 1];
443         yr = a[k1];
444         yi = a[k1 + 1];
445         a[j1] = yr;
446         a[j1 + 1] = yi;
447         a[k1] = xr;
448         a[k1 + 1] = xi;
449       }
450     }
451   }
452 }
453 
cftfsub(size_t n,float * a,float * w)454 void cftfsub(size_t n, float* a, float* w) {
455   size_t j, j1, j2, j3, l;
456   float x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i;
457 
458   l = 2;
459   if (n > 8) {
460     cft1st(n, a, w);
461     l = 8;
462     while ((l << 2) < n) {
463       cftmdl(n, l, a, w);
464       l <<= 2;
465     }
466   }
467   if ((l << 2) == n) {
468     for (j = 0; j < l; j += 2) {
469       j1 = j + l;
470       j2 = j1 + l;
471       j3 = j2 + l;
472       x0r = a[j] + a[j1];
473       x0i = a[j + 1] + a[j1 + 1];
474       x1r = a[j] - a[j1];
475       x1i = a[j + 1] - a[j1 + 1];
476       x2r = a[j2] + a[j3];
477       x2i = a[j2 + 1] + a[j3 + 1];
478       x3r = a[j2] - a[j3];
479       x3i = a[j2 + 1] - a[j3 + 1];
480       a[j] = x0r + x2r;
481       a[j + 1] = x0i + x2i;
482       a[j2] = x0r - x2r;
483       a[j2 + 1] = x0i - x2i;
484       a[j1] = x1r - x3i;
485       a[j1 + 1] = x1i + x3r;
486       a[j3] = x1r + x3i;
487       a[j3 + 1] = x1i - x3r;
488     }
489   } else {
490     for (j = 0; j < l; j += 2) {
491       j1 = j + l;
492       x0r = a[j] - a[j1];
493       x0i = a[j + 1] - a[j1 + 1];
494       a[j] += a[j1];
495       a[j + 1] += a[j1 + 1];
496       a[j1] = x0r;
497       a[j1 + 1] = x0i;
498     }
499   }
500 }
501 
cftbsub(size_t n,float * a,float * w)502 void cftbsub(size_t n, float* a, float* w) {
503   size_t j, j1, j2, j3, l;
504   float x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i;
505 
506   l = 2;
507   if (n > 8) {
508     cft1st(n, a, w);
509     l = 8;
510     while ((l << 2) < n) {
511       cftmdl(n, l, a, w);
512       l <<= 2;
513     }
514   }
515   if ((l << 2) == n) {
516     for (j = 0; j < l; j += 2) {
517       j1 = j + l;
518       j2 = j1 + l;
519       j3 = j2 + l;
520       x0r = a[j] + a[j1];
521       x0i = -a[j + 1] - a[j1 + 1];
522       x1r = a[j] - a[j1];
523       x1i = -a[j + 1] + a[j1 + 1];
524       x2r = a[j2] + a[j3];
525       x2i = a[j2 + 1] + a[j3 + 1];
526       x3r = a[j2] - a[j3];
527       x3i = a[j2 + 1] - a[j3 + 1];
528       a[j] = x0r + x2r;
529       a[j + 1] = x0i - x2i;
530       a[j2] = x0r - x2r;
531       a[j2 + 1] = x0i + x2i;
532       a[j1] = x1r - x3i;
533       a[j1 + 1] = x1i - x3r;
534       a[j3] = x1r + x3i;
535       a[j3 + 1] = x1i + x3r;
536     }
537   } else {
538     for (j = 0; j < l; j += 2) {
539       j1 = j + l;
540       x0r = a[j] - a[j1];
541       x0i = -a[j + 1] + a[j1 + 1];
542       a[j] += a[j1];
543       a[j + 1] = -a[j + 1] - a[j1 + 1];
544       a[j1] = x0r;
545       a[j1 + 1] = x0i;
546     }
547   }
548 }
549 
cft1st(size_t n,float * a,float * w)550 void cft1st(size_t n, float* a, float* w) {
551   size_t j, k1, k2;
552   float wk1r, wk1i, wk2r, wk2i, wk3r, wk3i;
553   float x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i;
554 
555   x0r = a[0] + a[2];
556   x0i = a[1] + a[3];
557   x1r = a[0] - a[2];
558   x1i = a[1] - a[3];
559   x2r = a[4] + a[6];
560   x2i = a[5] + a[7];
561   x3r = a[4] - a[6];
562   x3i = a[5] - a[7];
563   a[0] = x0r + x2r;
564   a[1] = x0i + x2i;
565   a[4] = x0r - x2r;
566   a[5] = x0i - x2i;
567   a[2] = x1r - x3i;
568   a[3] = x1i + x3r;
569   a[6] = x1r + x3i;
570   a[7] = x1i - x3r;
571   wk1r = w[2];
572   x0r = a[8] + a[10];
573   x0i = a[9] + a[11];
574   x1r = a[8] - a[10];
575   x1i = a[9] - a[11];
576   x2r = a[12] + a[14];
577   x2i = a[13] + a[15];
578   x3r = a[12] - a[14];
579   x3i = a[13] - a[15];
580   a[8] = x0r + x2r;
581   a[9] = x0i + x2i;
582   a[12] = x2i - x0i;
583   a[13] = x0r - x2r;
584   x0r = x1r - x3i;
585   x0i = x1i + x3r;
586   a[10] = wk1r * (x0r - x0i);
587   a[11] = wk1r * (x0r + x0i);
588   x0r = x3i + x1r;
589   x0i = x3r - x1i;
590   a[14] = wk1r * (x0i - x0r);
591   a[15] = wk1r * (x0i + x0r);
592   k1 = 0;
593   for (j = 16; j < n; j += 16) {
594     k1 += 2;
595     k2 = 2 * k1;
596     wk2r = w[k1];
597     wk2i = w[k1 + 1];
598     wk1r = w[k2];
599     wk1i = w[k2 + 1];
600     wk3r = wk1r - 2 * wk2i * wk1i;
601     wk3i = 2 * wk2i * wk1r - wk1i;
602     x0r = a[j] + a[j + 2];
603     x0i = a[j + 1] + a[j + 3];
604     x1r = a[j] - a[j + 2];
605     x1i = a[j + 1] - a[j + 3];
606     x2r = a[j + 4] + a[j + 6];
607     x2i = a[j + 5] + a[j + 7];
608     x3r = a[j + 4] - a[j + 6];
609     x3i = a[j + 5] - a[j + 7];
610     a[j] = x0r + x2r;
611     a[j + 1] = x0i + x2i;
612     x0r -= x2r;
613     x0i -= x2i;
614     a[j + 4] = wk2r * x0r - wk2i * x0i;
615     a[j + 5] = wk2r * x0i + wk2i * x0r;
616     x0r = x1r - x3i;
617     x0i = x1i + x3r;
618     a[j + 2] = wk1r * x0r - wk1i * x0i;
619     a[j + 3] = wk1r * x0i + wk1i * x0r;
620     x0r = x1r + x3i;
621     x0i = x1i - x3r;
622     a[j + 6] = wk3r * x0r - wk3i * x0i;
623     a[j + 7] = wk3r * x0i + wk3i * x0r;
624     wk1r = w[k2 + 2];
625     wk1i = w[k2 + 3];
626     wk3r = wk1r - 2 * wk2r * wk1i;
627     wk3i = 2 * wk2r * wk1r - wk1i;
628     x0r = a[j + 8] + a[j + 10];
629     x0i = a[j + 9] + a[j + 11];
630     x1r = a[j + 8] - a[j + 10];
631     x1i = a[j + 9] - a[j + 11];
632     x2r = a[j + 12] + a[j + 14];
633     x2i = a[j + 13] + a[j + 15];
634     x3r = a[j + 12] - a[j + 14];
635     x3i = a[j + 13] - a[j + 15];
636     a[j + 8] = x0r + x2r;
637     a[j + 9] = x0i + x2i;
638     x0r -= x2r;
639     x0i -= x2i;
640     a[j + 12] = -wk2i * x0r - wk2r * x0i;
641     a[j + 13] = -wk2i * x0i + wk2r * x0r;
642     x0r = x1r - x3i;
643     x0i = x1i + x3r;
644     a[j + 10] = wk1r * x0r - wk1i * x0i;
645     a[j + 11] = wk1r * x0i + wk1i * x0r;
646     x0r = x1r + x3i;
647     x0i = x1i - x3r;
648     a[j + 14] = wk3r * x0r - wk3i * x0i;
649     a[j + 15] = wk3r * x0i + wk3i * x0r;
650   }
651 }
652 
cftmdl(size_t n,size_t l,float * a,float * w)653 void cftmdl(size_t n, size_t l, float* a, float* w) {
654   size_t j, j1, j2, j3, k, k1, k2, m, m2;
655   float wk1r, wk1i, wk2r, wk2i, wk3r, wk3i;
656   float x0r, x0i, x1r, x1i, x2r, x2i, x3r, x3i;
657 
658   m = l << 2;
659   for (j = 0; j < l; j += 2) {
660     j1 = j + l;
661     j2 = j1 + l;
662     j3 = j2 + l;
663     x0r = a[j] + a[j1];
664     x0i = a[j + 1] + a[j1 + 1];
665     x1r = a[j] - a[j1];
666     x1i = a[j + 1] - a[j1 + 1];
667     x2r = a[j2] + a[j3];
668     x2i = a[j2 + 1] + a[j3 + 1];
669     x3r = a[j2] - a[j3];
670     x3i = a[j2 + 1] - a[j3 + 1];
671     a[j] = x0r + x2r;
672     a[j + 1] = x0i + x2i;
673     a[j2] = x0r - x2r;
674     a[j2 + 1] = x0i - x2i;
675     a[j1] = x1r - x3i;
676     a[j1 + 1] = x1i + x3r;
677     a[j3] = x1r + x3i;
678     a[j3 + 1] = x1i - x3r;
679   }
680   wk1r = w[2];
681   for (j = m; j < l + m; j += 2) {
682     j1 = j + l;
683     j2 = j1 + l;
684     j3 = j2 + l;
685     x0r = a[j] + a[j1];
686     x0i = a[j + 1] + a[j1 + 1];
687     x1r = a[j] - a[j1];
688     x1i = a[j + 1] - a[j1 + 1];
689     x2r = a[j2] + a[j3];
690     x2i = a[j2 + 1] + a[j3 + 1];
691     x3r = a[j2] - a[j3];
692     x3i = a[j2 + 1] - a[j3 + 1];
693     a[j] = x0r + x2r;
694     a[j + 1] = x0i + x2i;
695     a[j2] = x2i - x0i;
696     a[j2 + 1] = x0r - x2r;
697     x0r = x1r - x3i;
698     x0i = x1i + x3r;
699     a[j1] = wk1r * (x0r - x0i);
700     a[j1 + 1] = wk1r * (x0r + x0i);
701     x0r = x3i + x1r;
702     x0i = x3r - x1i;
703     a[j3] = wk1r * (x0i - x0r);
704     a[j3 + 1] = wk1r * (x0i + x0r);
705   }
706   k1 = 0;
707   m2 = 2 * m;
708   for (k = m2; k < n; k += m2) {
709     k1 += 2;
710     k2 = 2 * k1;
711     wk2r = w[k1];
712     wk2i = w[k1 + 1];
713     wk1r = w[k2];
714     wk1i = w[k2 + 1];
715     wk3r = wk1r - 2 * wk2i * wk1i;
716     wk3i = 2 * wk2i * wk1r - wk1i;
717     for (j = k; j < l + k; j += 2) {
718       j1 = j + l;
719       j2 = j1 + l;
720       j3 = j2 + l;
721       x0r = a[j] + a[j1];
722       x0i = a[j + 1] + a[j1 + 1];
723       x1r = a[j] - a[j1];
724       x1i = a[j + 1] - a[j1 + 1];
725       x2r = a[j2] + a[j3];
726       x2i = a[j2 + 1] + a[j3 + 1];
727       x3r = a[j2] - a[j3];
728       x3i = a[j2 + 1] - a[j3 + 1];
729       a[j] = x0r + x2r;
730       a[j + 1] = x0i + x2i;
731       x0r -= x2r;
732       x0i -= x2i;
733       a[j2] = wk2r * x0r - wk2i * x0i;
734       a[j2 + 1] = wk2r * x0i + wk2i * x0r;
735       x0r = x1r - x3i;
736       x0i = x1i + x3r;
737       a[j1] = wk1r * x0r - wk1i * x0i;
738       a[j1 + 1] = wk1r * x0i + wk1i * x0r;
739       x0r = x1r + x3i;
740       x0i = x1i - x3r;
741       a[j3] = wk3r * x0r - wk3i * x0i;
742       a[j3 + 1] = wk3r * x0i + wk3i * x0r;
743     }
744     wk1r = w[k2 + 2];
745     wk1i = w[k2 + 3];
746     wk3r = wk1r - 2 * wk2r * wk1i;
747     wk3i = 2 * wk2r * wk1r - wk1i;
748     for (j = k + m; j < l + (k + m); j += 2) {
749       j1 = j + l;
750       j2 = j1 + l;
751       j3 = j2 + l;
752       x0r = a[j] + a[j1];
753       x0i = a[j + 1] + a[j1 + 1];
754       x1r = a[j] - a[j1];
755       x1i = a[j + 1] - a[j1 + 1];
756       x2r = a[j2] + a[j3];
757       x2i = a[j2 + 1] + a[j3 + 1];
758       x3r = a[j2] - a[j3];
759       x3i = a[j2 + 1] - a[j3 + 1];
760       a[j] = x0r + x2r;
761       a[j + 1] = x0i + x2i;
762       x0r -= x2r;
763       x0i -= x2i;
764       a[j2] = -wk2i * x0r - wk2r * x0i;
765       a[j2 + 1] = -wk2i * x0i + wk2r * x0r;
766       x0r = x1r - x3i;
767       x0i = x1i + x3r;
768       a[j1] = wk1r * x0r - wk1i * x0i;
769       a[j1 + 1] = wk1r * x0i + wk1i * x0r;
770       x0r = x1r + x3i;
771       x0i = x1i - x3r;
772       a[j3] = wk3r * x0r - wk3i * x0i;
773       a[j3 + 1] = wk3r * x0i + wk3i * x0r;
774     }
775   }
776 }
777 
rftfsub(size_t n,float * a,size_t nc,float * c)778 void rftfsub(size_t n, float* a, size_t nc, float* c) {
779   size_t j, k, kk, ks, m;
780   float wkr, wki, xr, xi, yr, yi;
781 
782   m = n >> 1;
783   ks = 2 * nc / m;
784   kk = 0;
785   for (j = 2; j < m; j += 2) {
786     k = n - j;
787     kk += ks;
788     wkr = 0.5f - c[nc - kk];
789     wki = c[kk];
790     xr = a[j] - a[k];
791     xi = a[j + 1] + a[k + 1];
792     yr = wkr * xr - wki * xi;
793     yi = wkr * xi + wki * xr;
794     a[j] -= yr;
795     a[j + 1] -= yi;
796     a[k] += yr;
797     a[k + 1] -= yi;
798   }
799 }
800 
rftbsub(size_t n,float * a,size_t nc,float * c)801 void rftbsub(size_t n, float* a, size_t nc, float* c) {
802   size_t j, k, kk, ks, m;
803   float wkr, wki, xr, xi, yr, yi;
804 
805   a[1] = -a[1];
806   m = n >> 1;
807   ks = 2 * nc / m;
808   kk = 0;
809   for (j = 2; j < m; j += 2) {
810     k = n - j;
811     kk += ks;
812     wkr = 0.5f - c[nc - kk];
813     wki = c[kk];
814     xr = a[j] - a[k];
815     xi = a[j + 1] + a[k + 1];
816     yr = wkr * xr + wki * xi;
817     yi = wkr * xi - wki * xr;
818     a[j] -= yr;
819     a[j + 1] = yi - a[j + 1];
820     a[k] += yr;
821     a[k + 1] = yi - a[k + 1];
822   }
823   a[m + 1] = -a[m + 1];
824 }
825 
826 }  // namespace
827 
WebRtc_rdft(size_t n,int isgn,float * a,size_t * ip,float * w)828 void WebRtc_rdft(size_t n, int isgn, float* a, size_t* ip, float* w) {
829   size_t nw, nc;
830   float xi;
831 
832   nw = ip[0];
833   if (n > (nw << 2)) {
834     nw = n >> 2;
835     makewt(nw, ip, w);
836   }
837   nc = ip[1];
838   if (n > (nc << 2)) {
839     nc = n >> 2;
840     makect(nc, ip, w + nw);
841   }
842   if (isgn >= 0) {
843     if (n > 4) {
844       bitrv2(n, ip + 2, a);
845       cftfsub(n, a, w);
846       rftfsub(n, a, nc, w + nw);
847     } else if (n == 4) {
848       cftfsub(n, a, w);
849     }
850     xi = a[0] - a[1];
851     a[0] += a[1];
852     a[1] = xi;
853   } else {
854     a[1] = 0.5f * (a[0] - a[1]);
855     a[0] -= a[1];
856     if (n > 4) {
857       rftbsub(n, a, nc, w + nw);
858       bitrv2(n, ip + 2, a);
859       cftbsub(n, a, w);
860     } else if (n == 4) {
861       cftfsub(n, a, w);
862     }
863   }
864 }
865 
866 }  // namespace webrtc
867