2 Copyright (C) 2003-2009 Paul Brossier <piem@aubio.org>
4 This file is part of aubio.
6 aubio is free software: you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation, either version 3 of the License, or
9 (at your option) any later version.
11 aubio is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
16 You should have received a copy of the GNU General Public License
17 along with aubio. If not, see <http://www.gnu.org/licenses/>.
21 #include "aubio_priv.h"
24 #include "mathutils.h"
25 #include "spectral/fft.h"
27 #ifdef HAVE_FFTW3 // using FFTW3
28 /* note that <complex.h> is not included here but only in aubio_priv.h, so that
29 * c++ projects can still use their own complex definition. */
35 /** fft data type with complex.h and fftw3f */
36 #define FFTW_TYPE fftwf_complex
38 /** fft data type with complex.h and fftw3 */
39 #define FFTW_TYPE fftw_complex
43 /** fft data type without complex.h and with fftw3f */
44 #define FFTW_TYPE float
46 /** fft data type without complex.h and with fftw */
47 #define FFTW_TYPE double
52 typedef FFTW_TYPE fft_data_t;
55 #define fftw_malloc fftwf_malloc
56 #define fftw_free fftwf_free
57 #define fftw_execute fftwf_execute
58 #define fftw_plan_dft_r2c_1d fftwf_plan_dft_r2c_1d
59 #define fftw_plan_dft_c2r_1d fftwf_plan_dft_c2r_1d
60 #define fftw_plan_r2r_1d fftwf_plan_r2r_1d
61 #define fftw_plan fftwf_plan
62 #define fftw_destroy_plan fftwf_destroy_plan
67 #error "Using aubio in double precision with fftw3 in single precision"
68 #endif /* HAVE_AUBIO_DOUBLE */
70 #elif defined (HAVE_FFTW3) /* HAVE_FFTW3F */
71 #if !HAVE_AUBIO_DOUBLE
72 #error "Using aubio in single precision with fftw3 in double precision"
73 #endif /* HAVE_AUBIO_DOUBLE */
75 #endif /* HAVE_FFTW3F */
77 // a global mutex for FFTW thread safety
78 pthread_mutex_t aubio_fftw_mutex = PTHREAD_MUTEX_INITIALIZER;
80 #elif defined HAVE_ACCELERATE // using ACCELERATE
81 // https://developer.apple.com/library/mac/#documentation/Accelerate/Reference/vDSPRef/Reference/reference.html
82 #include <Accelerate/Accelerate.h>
84 #if !HAVE_AUBIO_DOUBLE
85 #define aubio_vDSP_ctoz vDSP_ctoz
86 #define aubio_vDSP_fft_zrip vDSP_fft_zrip
87 #define aubio_vDSP_ztoc vDSP_ztoc
88 #define aubio_vDSP_zvmags vDSP_zvmags
89 #define aubio_vDSP_zvphas vDSP_zvphas
90 #define aubio_vDSP_vsadd vDSP_vsadd
91 #define aubio_vDSP_vsmul vDSP_vsmul
92 #define aubio_DSPComplex DSPComplex
93 #define aubio_DSPSplitComplex DSPSplitComplex
94 #define aubio_vDSP_DFT_Setup vDSP_DFT_Setup
95 #define aubio_vDSP_DFT_zrop_CreateSetup vDSP_DFT_zrop_CreateSetup
96 #define aubio_vDSP_DFT_Execute vDSP_DFT_Execute
97 #define aubio_vDSP_DFT_DestroySetup vDSP_DFT_DestroySetup
98 #define aubio_vvsqrt vvsqrtf
100 #define aubio_vDSP_ctoz vDSP_ctozD
101 #define aubio_vDSP_fft_zrip vDSP_fft_zripD
102 #define aubio_vDSP_ztoc vDSP_ztocD
103 #define aubio_vDSP_zvmags vDSP_zvmagsD
104 #define aubio_vDSP_zvphas vDSP_zvphasD
105 #define aubio_vDSP_vsadd vDSP_vsaddD
106 #define aubio_vDSP_vsmul vDSP_vsmulD
107 #define aubio_DSPComplex DSPDoubleComplex
108 #define aubio_DSPSplitComplex DSPDoubleSplitComplex
109 #define aubio_vDSP_DFT_Setup vDSP_DFT_SetupD
110 #define aubio_vDSP_DFT_zrop_CreateSetup vDSP_DFT_zrop_CreateSetupD
111 #define aubio_vDSP_DFT_Execute vDSP_DFT_ExecuteD
112 #define aubio_vDSP_DFT_DestroySetup vDSP_DFT_DestroySetupD
113 #define aubio_vvsqrt vvsqrt
114 #endif /* HAVE_AUBIO_DOUBLE */
116 #elif defined HAVE_INTEL_IPP // using INTEL IPP
118 #if !HAVE_AUBIO_DOUBLE
119 #define aubio_IppFloat Ipp32f
120 #define aubio_IppComplex Ipp32fc
121 #define aubio_FFTSpec FFTSpec_R_32f
122 #define aubio_ippsMalloc_complex ippsMalloc_32fc
123 #define aubio_ippsFFTInit_R ippsFFTInit_R_32f
124 #define aubio_ippsFFTGetSize_R ippsFFTGetSize_R_32f
125 #define aubio_ippsFFTInv_CCSToR ippsFFTInv_CCSToR_32f
126 #define aubio_ippsFFTFwd_RToCCS ippsFFTFwd_RToCCS_32f
127 #define aubio_ippsAtan2 ippsAtan2_32f_A21
128 #else /* HAVE_AUBIO_DOUBLE */
129 #define aubio_IppFloat Ipp64f
130 #define aubio_IppComplex Ipp64fc
131 #define aubio_FFTSpec FFTSpec_R_64f
132 #define aubio_ippsMalloc_complex ippsMalloc_64fc
133 #define aubio_ippsFFTInit_R ippsFFTInit_R_64f
134 #define aubio_ippsFFTGetSize_R ippsFFTGetSize_R_64f
135 #define aubio_ippsFFTInv_CCSToR ippsFFTInv_CCSToR_64f
136 #define aubio_ippsFFTFwd_RToCCS ippsFFTFwd_RToCCS_64f
137 #define aubio_ippsAtan2 ippsAtan2_64f_A50
142 // let's use ooura instead
143 extern void aubio_ooura_rdft(int, int, smpl_t *, int *, smpl_t *);
147 struct _aubio_fft_t {
151 #ifdef HAVE_FFTW3 // using FFTW3
154 fft_data_t * specdata; /* complex spectral data */
156 #elif defined HAVE_ACCELERATE // using ACCELERATE
157 aubio_vDSP_DFT_Setup fftSetupFwd;
158 aubio_vDSP_DFT_Setup fftSetupBwd;
159 aubio_DSPSplitComplex spec;
162 #elif defined HAVE_INTEL_IPP // using Intel IPP
167 struct aubio_FFTSpec* fftSpec;
168 aubio_IppComplex* complexOut;
173 #endif /* using OOURA */
178 aubio_fft_t * new_aubio_fft (uint_t winsize) {
179 aubio_fft_t * s = AUBIO_NEW(aubio_fft_t);
180 if ((sint_t)winsize < 2) {
181 AUBIO_ERR("fft: got winsize %d, but can not be < 2\n", winsize);
187 s->winsize = winsize;
188 /* allocate memory */
189 s->in = AUBIO_ARRAY(real_t,winsize);
190 s->out = AUBIO_ARRAY(real_t,winsize);
191 s->compspec = new_fvec(winsize);
193 pthread_mutex_lock(&aubio_fftw_mutex);
194 #ifdef HAVE_COMPLEX_H
195 s->fft_size = winsize/2 + 1;
196 s->specdata = (fft_data_t*)fftw_malloc(sizeof(fft_data_t)*s->fft_size);
197 s->pfw = fftw_plan_dft_r2c_1d(winsize, s->in, s->specdata, FFTW_ESTIMATE);
198 s->pbw = fftw_plan_dft_c2r_1d(winsize, s->specdata, s->out, FFTW_ESTIMATE);
200 s->fft_size = winsize;
201 s->specdata = (fft_data_t*)fftw_malloc(sizeof(fft_data_t)*s->fft_size);
202 s->pfw = fftw_plan_r2r_1d(winsize, s->in, s->specdata, FFTW_R2HC, FFTW_ESTIMATE);
203 s->pbw = fftw_plan_r2r_1d(winsize, s->specdata, s->out, FFTW_HC2R, FFTW_ESTIMATE);
205 pthread_mutex_unlock(&aubio_fftw_mutex);
206 for (i = 0; i < s->winsize; i++) {
210 for (i = 0; i < s->fft_size; i++) {
214 #elif defined HAVE_ACCELERATE // using ACCELERATE
216 uint_t radix = winsize;
218 while ((radix / 2) * 2 == radix) {
222 if (order < 4 || (radix != 1 && radix != 3 && radix != 5 && radix != 15)) {
223 AUBIO_ERR("fft: vDSP/Accelerate supports FFT with sizes = "
224 "f * 2 ** n, where n > 4 and f in [1, 3, 5, 15], but requested %d. "
225 "Use the closest power of two, or try recompiling aubio with "
226 "--enable-fftw3.\n", winsize);
230 s->winsize = winsize;
231 s->fft_size = winsize;
232 s->compspec = new_fvec(winsize);
233 s->in = AUBIO_ARRAY(smpl_t, s->fft_size);
234 s->out = AUBIO_ARRAY(smpl_t, s->fft_size);
235 s->spec.realp = AUBIO_ARRAY(smpl_t, s->fft_size/2);
236 s->spec.imagp = AUBIO_ARRAY(smpl_t, s->fft_size/2);
237 s->fftSetupFwd = aubio_vDSP_DFT_zrop_CreateSetup(NULL,
238 s->fft_size, vDSP_DFT_FORWARD);
239 s->fftSetupBwd = aubio_vDSP_DFT_zrop_CreateSetup(s->fftSetupFwd,
240 s->fft_size, vDSP_DFT_INVERSE);
242 #elif defined HAVE_INTEL_IPP // using Intel IPP
243 const IppHintAlgorithm qualityHint = ippAlgHintAccurate; // OR ippAlgHintFast;
244 const int flags = IPP_FFT_NODIV_BY_ANY; // we're scaling manually afterwards
245 int order = aubio_power_of_two_order(winsize);
246 int sizeSpec, sizeInit, sizeBuffer;
249 if (winsize <= 4 || aubio_is_power_of_two(winsize) != 1)
251 AUBIO_ERR("intel IPP fft: can only create with sizes > 4 and power of two, requested %d,"
252 " try recompiling aubio with --enable-fftw3\n", winsize);
256 status = aubio_ippsFFTGetSize_R(order, flags, qualityHint,
257 &sizeSpec, &sizeInit, &sizeBuffer);
258 if (status != ippStsNoErr) {
259 AUBIO_ERR("fft: failed to initialize fft. IPP error: %d\n", status);
262 s->fft_size = s->winsize = winsize;
263 s->compspec = new_fvec(winsize);
264 s->in = AUBIO_ARRAY(smpl_t, s->winsize);
265 s->out = AUBIO_ARRAY(smpl_t, s->winsize);
266 s->memSpec = ippsMalloc_8u(sizeSpec);
267 s->memBuffer = ippsMalloc_8u(sizeBuffer);
269 s->memInit = ippsMalloc_8u(sizeInit);
271 s->complexOut = aubio_ippsMalloc_complex(s->fft_size / 2 + 1);
272 status = aubio_ippsFFTInit_R(
273 &s->fftSpec, order, flags, qualityHint, s->memSpec, s->memInit);
274 if (status != ippStsNoErr) {
275 AUBIO_ERR("fft: failed to initialize. IPP error: %d\n", status);
280 if (aubio_is_power_of_two(winsize) != 1) {
281 AUBIO_ERR("fft: can only create with sizes power of two, requested %d,"
282 " try recompiling aubio with --enable-fftw3\n", winsize);
285 s->winsize = winsize;
286 s->fft_size = winsize / 2 + 1;
287 s->compspec = new_fvec(winsize);
288 s->in = AUBIO_ARRAY(smpl_t, s->winsize);
289 s->out = AUBIO_ARRAY(smpl_t, s->winsize);
290 s->ip = AUBIO_ARRAY(int , s->fft_size);
291 s->w = AUBIO_ARRAY(smpl_t, s->fft_size);
293 #endif /* using OOURA */
302 void del_aubio_fft(aubio_fft_t * s) {
304 #ifdef HAVE_FFTW3 // using FFTW3
305 pthread_mutex_lock(&aubio_fftw_mutex);
306 fftw_destroy_plan(s->pfw);
307 fftw_destroy_plan(s->pbw);
308 fftw_free(s->specdata);
309 pthread_mutex_unlock(&aubio_fftw_mutex);
311 #elif defined HAVE_ACCELERATE // using ACCELERATE
312 AUBIO_FREE(s->spec.realp);
313 AUBIO_FREE(s->spec.imagp);
314 aubio_vDSP_DFT_DestroySetup(s->fftSetupBwd);
315 aubio_vDSP_DFT_DestroySetup(s->fftSetupFwd);
317 #elif defined HAVE_INTEL_IPP // using Intel IPP
320 ippFree(s->memBuffer);
321 ippFree(s->complexOut);
328 del_fvec(s->compspec);
334 void aubio_fft_do(aubio_fft_t * s, const fvec_t * input, cvec_t * spectrum) {
335 aubio_fft_do_complex(s, input, s->compspec);
336 aubio_fft_get_spectrum(s->compspec, spectrum);
339 void aubio_fft_rdo(aubio_fft_t * s, const cvec_t * spectrum, fvec_t * output) {
340 aubio_fft_get_realimag(spectrum, s->compspec);
341 aubio_fft_rdo_complex(s, s->compspec, output);
344 void aubio_fft_do_complex(aubio_fft_t * s, const fvec_t * input, fvec_t * compspec) {
346 #ifndef HAVE_MEMCPY_HACKS
347 for (i=0; i < s->winsize; i++) {
348 s->in[i] = input->data[i];
351 memcpy(s->in, input->data, s->winsize * sizeof(smpl_t));
352 #endif /* HAVE_MEMCPY_HACKS */
354 #ifdef HAVE_FFTW3 // using FFTW3
355 fftw_execute(s->pfw);
356 #ifdef HAVE_COMPLEX_H
357 compspec->data[0] = REAL(s->specdata[0]);
358 for (i = 1; i < s->fft_size -1 ; i++) {
359 compspec->data[i] = REAL(s->specdata[i]);
360 compspec->data[compspec->length - i] = IMAG(s->specdata[i]);
362 compspec->data[s->fft_size-1] = REAL(s->specdata[s->fft_size-1]);
363 #else /* HAVE_COMPLEX_H */
364 for (i = 0; i < s->fft_size; i++) {
365 compspec->data[i] = s->specdata[i];
367 #endif /* HAVE_COMPLEX_H */
369 #elif defined HAVE_ACCELERATE // using ACCELERATE
370 // convert real data to even/odd format used in vDSP
371 aubio_vDSP_ctoz((aubio_DSPComplex*)s->in, 2, &s->spec, 1, s->fft_size/2);
373 aubio_vDSP_DFT_Execute(s->fftSetupFwd, s->spec.realp, s->spec.imagp,
374 s->spec.realp, s->spec.imagp);
375 // convert from vDSP complex split to [ r0, r1, ..., rN, iN-1, .., i2, i1]
376 compspec->data[0] = s->spec.realp[0];
377 compspec->data[s->fft_size / 2] = s->spec.imagp[0];
378 for (i = 1; i < s->fft_size / 2; i++) {
379 compspec->data[i] = s->spec.realp[i];
380 compspec->data[s->fft_size - i] = s->spec.imagp[i];
383 smpl_t scale = 1./2.;
384 aubio_vDSP_vsmul(compspec->data, 1, &scale, compspec->data, 1, s->fft_size);
386 #elif defined HAVE_INTEL_IPP // using Intel IPP
389 aubio_ippsFFTFwd_RToCCS(s->in, (aubio_IppFloat*)s->complexOut, s->fftSpec, s->memBuffer);
390 // convert complex buffer to [ r0, r1, ..., rN, iN-1, .., i2, i1]
391 compspec->data[0] = s->complexOut[0].re;
392 compspec->data[s->fft_size / 2] = s->complexOut[s->fft_size / 2].re;
393 for (i = 1; i < s->fft_size / 2; i++) {
394 compspec->data[i] = s->complexOut[i].re;
395 compspec->data[s->fft_size - i] = s->complexOut[i].im;
399 aubio_ooura_rdft(s->winsize, 1, s->in, s->ip, s->w);
400 compspec->data[0] = s->in[0];
401 compspec->data[s->winsize / 2] = s->in[1];
402 for (i = 1; i < s->fft_size - 1; i++) {
403 compspec->data[i] = s->in[2 * i];
404 compspec->data[s->winsize - i] = - s->in[2 * i + 1];
406 #endif /* using OOURA */
409 void aubio_fft_rdo_complex(aubio_fft_t * s, const fvec_t * compspec, fvec_t * output) {
412 const smpl_t renorm = 1./(smpl_t)s->winsize;
413 #ifdef HAVE_COMPLEX_H
414 s->specdata[0] = compspec->data[0];
415 for (i=1; i < s->fft_size - 1; i++) {
416 s->specdata[i] = compspec->data[i] +
417 I * compspec->data[compspec->length - i];
419 s->specdata[s->fft_size - 1] = compspec->data[s->fft_size - 1];
421 for (i=0; i < s->fft_size; i++) {
422 s->specdata[i] = compspec->data[i];
425 fftw_execute(s->pbw);
426 for (i = 0; i < output->length; i++) {
427 output->data[i] = s->out[i]*renorm;
430 #elif defined HAVE_ACCELERATE // using ACCELERATE
431 // convert from real imag [ r0, r1, ..., rN, iN-1, .., i2, i1]
432 // to vDSP packed format [ r0, rN, r1, i1, ..., rN-1, iN-1 ]
433 s->out[0] = compspec->data[0];
434 s->out[1] = compspec->data[s->winsize / 2];
435 for (i = 1; i < s->fft_size / 2; i++) {
436 s->out[2 * i] = compspec->data[i];
437 s->out[2 * i + 1] = compspec->data[s->winsize - i];
439 // convert to split complex format used in vDSP
440 aubio_vDSP_ctoz((aubio_DSPComplex*)s->out, 2, &s->spec, 1, s->fft_size/2);
442 aubio_vDSP_DFT_Execute(s->fftSetupBwd, s->spec.realp, s->spec.imagp,
443 s->spec.realp, s->spec.imagp);
444 // convert result to real output
445 aubio_vDSP_ztoc(&s->spec, 1, (aubio_DSPComplex*)output->data, 2, s->fft_size/2);
447 smpl_t scale = 1.0 / s->winsize;
448 aubio_vDSP_vsmul(output->data, 1, &scale, output->data, 1, s->fft_size);
450 #elif defined HAVE_INTEL_IPP // using Intel IPP
452 // convert from real imag [ r0, 0, ..., rN, iN-1, .., i2, i1, iN-1] to complex format
453 s->complexOut[0].re = compspec->data[0];
454 s->complexOut[0].im = 0;
455 s->complexOut[s->fft_size / 2].re = compspec->data[s->fft_size / 2];
456 s->complexOut[s->fft_size / 2].im = 0.0;
457 for (i = 1; i < s->fft_size / 2; i++) {
458 s->complexOut[i].re = compspec->data[i];
459 s->complexOut[i].im = compspec->data[s->fft_size - i];
462 aubio_ippsFFTInv_CCSToR((const aubio_IppFloat *)s->complexOut, output->data, s->fftSpec, s->memBuffer);
464 aubio_ippsMulC(output->data, 1.0 / s->winsize, output->data, s->fft_size);
467 smpl_t scale = 2.0 / s->winsize;
468 s->out[0] = compspec->data[0];
469 s->out[1] = compspec->data[s->winsize / 2];
470 for (i = 1; i < s->fft_size - 1; i++) {
471 s->out[2 * i] = compspec->data[i];
472 s->out[2 * i + 1] = - compspec->data[s->winsize - i];
474 aubio_ooura_rdft(s->winsize, -1, s->out, s->ip, s->w);
475 for (i=0; i < s->winsize; i++) {
476 output->data[i] = s->out[i] * scale;
481 void aubio_fft_get_spectrum(const fvec_t * compspec, cvec_t * spectrum) {
482 aubio_fft_get_phas(compspec, spectrum);
483 aubio_fft_get_norm(compspec, spectrum);
486 void aubio_fft_get_realimag(const cvec_t * spectrum, fvec_t * compspec) {
487 aubio_fft_get_imag(spectrum, compspec);
488 aubio_fft_get_real(spectrum, compspec);
491 void aubio_fft_get_phas(const fvec_t * compspec, cvec_t * spectrum) {
493 if (compspec->data[0] < 0) {
494 spectrum->phas[0] = PI;
496 spectrum->phas[0] = 0.;
498 #if defined(HAVE_INTEL_IPP)
499 // convert from real imag [ r0, r1, ..., rN, iN-1, ..., i2, i1, i0]
500 // to [ r0, r1, ..., rN, i0, i1, i2, ..., iN-1]
501 for (i = 1; i < spectrum->length / 2; i++) {
502 ELEM_SWAP(compspec->data[compspec->length - i],
503 compspec->data[spectrum->length + i - 1]);
505 aubio_ippsAtan2(compspec->data + spectrum->length,
506 compspec->data + 1, spectrum->phas + 1, spectrum->length - 1);
507 // revert the imaginary part back again
508 for (i = 1; i < spectrum->length / 2; i++) {
509 ELEM_SWAP(compspec->data[spectrum->length + i - 1],
510 compspec->data[compspec->length - i]);
513 for (i=1; i < spectrum->length - 1; i++) {
514 spectrum->phas[i] = ATAN2(compspec->data[compspec->length-i],
519 // for even length only, make sure last element is 0 or PI
520 if (2 * (compspec->length / 2) == compspec->length) {
522 if (compspec->data[compspec->length/2] < 0) {
523 spectrum->phas[spectrum->length - 1] = PI;
525 spectrum->phas[spectrum->length - 1] = 0.;
529 i = spectrum->length - 1;
530 spectrum->phas[i] = ATAN2(compspec->data[compspec->length-i],
536 void aubio_fft_get_norm(const fvec_t * compspec, cvec_t * spectrum) {
538 spectrum->norm[0] = ABS(compspec->data[0]);
539 for (i=1; i < spectrum->length - 1; i++) {
540 spectrum->norm[i] = SQRT(SQR(compspec->data[i])
541 + SQR(compspec->data[compspec->length - i]) );
544 // for even length, make sure last element is > 0
545 if (2 * (compspec->length / 2) == compspec->length) {
547 spectrum->norm[spectrum->length-1] =
548 ABS(compspec->data[compspec->length/2]);
551 i = spectrum->length - 1;
552 spectrum->norm[i] = SQRT(SQR(compspec->data[i])
553 + SQR(compspec->data[compspec->length - i]) );
558 void aubio_fft_get_imag(const cvec_t * spectrum, fvec_t * compspec) {
560 for (i = 1; i < ( compspec->length + 1 ) / 2 /*- 1 + 1*/; i++) {
561 compspec->data[compspec->length - i] =
562 spectrum->norm[i]*SIN(spectrum->phas[i]);
566 void aubio_fft_get_real(const cvec_t * spectrum, fvec_t * compspec) {
568 for (i = 0; i < compspec->length / 2 + 1; i++) {
570 spectrum->norm[i]*COS(spectrum->phas[i]);