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_vDSP_create_fftsetup vDSP_create_fftsetup
93 #define aubio_vDSP_destroy_fftsetup vDSP_destroy_fftsetup
94 #define aubio_DSPComplex DSPComplex
95 #define aubio_DSPSplitComplex DSPSplitComplex
96 #define aubio_FFTSetup FFTSetup
97 #define aubio_vvsqrt vvsqrtf
99 #define aubio_vDSP_ctoz vDSP_ctozD
100 #define aubio_vDSP_fft_zrip vDSP_fft_zripD
101 #define aubio_vDSP_ztoc vDSP_ztocD
102 #define aubio_vDSP_zvmags vDSP_zvmagsD
103 #define aubio_vDSP_zvphas vDSP_zvphasD
104 #define aubio_vDSP_vsadd vDSP_vsaddD
105 #define aubio_vDSP_vsmul vDSP_vsmulD
106 #define aubio_vDSP_create_fftsetup vDSP_create_fftsetupD
107 #define aubio_vDSP_destroy_fftsetup vDSP_destroy_fftsetupD
108 #define aubio_DSPComplex DSPDoubleComplex
109 #define aubio_DSPSplitComplex DSPDoubleSplitComplex
110 #define aubio_FFTSetup FFTSetupD
111 #define aubio_vvsqrt vvsqrt
112 #endif /* HAVE_AUBIO_DOUBLE */
114 #elif defined HAVE_INTEL_IPP // using INTEL IPP
116 #if !HAVE_AUBIO_DOUBLE
117 #define aubio_IppFloat Ipp32f
118 #define aubio_IppComplex Ipp32fc
119 #define aubio_FFTSpec FFTSpec_R_32f
120 #define aubio_ippsMalloc_complex ippsMalloc_32fc
121 #define aubio_ippsFFTInit_R ippsFFTInit_R_32f
122 #define aubio_ippsFFTGetSize_R ippsFFTGetSize_R_32f
123 #define aubio_ippsFFTInv_CCSToR ippsFFTInv_CCSToR_32f
124 #define aubio_ippsFFTFwd_RToCCS ippsFFTFwd_RToCCS_32f
125 #define aubio_ippsAtan2 ippsAtan2_32f_A21
126 #else /* HAVE_AUBIO_DOUBLE */
127 #define aubio_IppFloat Ipp64f
128 #define aubio_IppComplex Ipp64fc
129 #define aubio_FFTSpec FFTSpec_R_64f
130 #define aubio_ippsMalloc_complex ippsMalloc_64fc
131 #define aubio_ippsFFTInit_R ippsFFTInit_R_64f
132 #define aubio_ippsFFTGetSize_R ippsFFTGetSize_R_64f
133 #define aubio_ippsFFTInv_CCSToR ippsFFTInv_CCSToR_64f
134 #define aubio_ippsFFTFwd_RToCCS ippsFFTFwd_RToCCS_64f
135 #define aubio_ippsAtan2 ippsAtan2_64f_A50
140 // let's use ooura instead
141 extern void aubio_ooura_rdft(int, int, smpl_t *, int *, smpl_t *);
145 struct _aubio_fft_t {
149 #ifdef HAVE_FFTW3 // using FFTW3
152 fft_data_t * specdata; /* complex spectral data */
154 #elif defined HAVE_ACCELERATE // using ACCELERATE
156 aubio_FFTSetup fftSetup;
157 aubio_DSPSplitComplex spec;
160 #elif defined HAVE_INTEL_IPP // using Intel IPP
165 struct aubio_FFTSpec* fftSpec;
166 aubio_IppComplex* complexOut;
171 #endif /* using OOURA */
176 aubio_fft_t * new_aubio_fft (uint_t winsize) {
177 aubio_fft_t * s = AUBIO_NEW(aubio_fft_t);
178 if ((sint_t)winsize < 2) {
179 AUBIO_ERR("fft: got winsize %d, but can not be < 2\n", winsize);
185 s->winsize = winsize;
186 /* allocate memory */
187 s->in = AUBIO_ARRAY(real_t,winsize);
188 s->out = AUBIO_ARRAY(real_t,winsize);
189 s->compspec = new_fvec(winsize);
191 pthread_mutex_lock(&aubio_fftw_mutex);
192 #ifdef HAVE_COMPLEX_H
193 s->fft_size = winsize/2 + 1;
194 s->specdata = (fft_data_t*)fftw_malloc(sizeof(fft_data_t)*s->fft_size);
195 s->pfw = fftw_plan_dft_r2c_1d(winsize, s->in, s->specdata, FFTW_ESTIMATE);
196 s->pbw = fftw_plan_dft_c2r_1d(winsize, s->specdata, s->out, FFTW_ESTIMATE);
198 s->fft_size = winsize;
199 s->specdata = (fft_data_t*)fftw_malloc(sizeof(fft_data_t)*s->fft_size);
200 s->pfw = fftw_plan_r2r_1d(winsize, s->in, s->specdata, FFTW_R2HC, FFTW_ESTIMATE);
201 s->pbw = fftw_plan_r2r_1d(winsize, s->specdata, s->out, FFTW_HC2R, FFTW_ESTIMATE);
203 pthread_mutex_unlock(&aubio_fftw_mutex);
204 for (i = 0; i < s->winsize; i++) {
208 for (i = 0; i < s->fft_size; i++) {
212 #elif defined HAVE_ACCELERATE // using ACCELERATE
213 s->winsize = winsize;
214 s->fft_size = winsize;
215 s->compspec = new_fvec(winsize);
216 s->log2fftsize = aubio_power_of_two_order(s->fft_size);
217 s->in = AUBIO_ARRAY(smpl_t, s->fft_size);
218 s->out = AUBIO_ARRAY(smpl_t, s->fft_size);
219 s->spec.realp = AUBIO_ARRAY(smpl_t, s->fft_size/2);
220 s->spec.imagp = AUBIO_ARRAY(smpl_t, s->fft_size/2);
221 s->fftSetup = aubio_vDSP_create_fftsetup(s->log2fftsize, FFT_RADIX2);
223 #elif defined HAVE_INTEL_IPP // using Intel IPP
224 const IppHintAlgorithm qualityHint = ippAlgHintAccurate; // OR ippAlgHintFast;
225 const int flags = IPP_FFT_NODIV_BY_ANY; // we're scaling manually afterwards
226 int order = aubio_power_of_two_order(winsize);
227 int sizeSpec, sizeInit, sizeBuffer;
230 if (winsize <= 4 || aubio_is_power_of_two(winsize) != 1)
232 AUBIO_ERR("intel IPP fft: can only create with sizes > 4 and power of two, requested %d,"
233 " try recompiling aubio with --enable-fftw3\n", winsize);
237 status = aubio_ippsFFTGetSize_R(order, flags, qualityHint,
238 &sizeSpec, &sizeInit, &sizeBuffer);
239 if (status != ippStsNoErr) {
240 AUBIO_ERR("fft: failed to initialize fft. IPP error: %d\n", status);
243 s->fft_size = s->winsize = winsize;
244 s->compspec = new_fvec(winsize);
245 s->in = AUBIO_ARRAY(smpl_t, s->winsize);
246 s->out = AUBIO_ARRAY(smpl_t, s->winsize);
247 s->memSpec = ippsMalloc_8u(sizeSpec);
248 s->memBuffer = ippsMalloc_8u(sizeBuffer);
250 s->memInit = ippsMalloc_8u(sizeInit);
252 s->complexOut = aubio_ippsMalloc_complex(s->fft_size / 2 + 1);
253 status = aubio_ippsFFTInit_R(
254 &s->fftSpec, order, flags, qualityHint, s->memSpec, s->memInit);
255 if (status != ippStsNoErr) {
256 AUBIO_ERR("fft: failed to initialize. IPP error: %d\n", status);
261 if (aubio_is_power_of_two(winsize) != 1) {
262 AUBIO_ERR("fft: can only create with sizes power of two, requested %d,"
263 " try recompiling aubio with --enable-fftw3\n", winsize);
266 s->winsize = winsize;
267 s->fft_size = winsize / 2 + 1;
268 s->compspec = new_fvec(winsize);
269 s->in = AUBIO_ARRAY(smpl_t, s->winsize);
270 s->out = AUBIO_ARRAY(smpl_t, s->winsize);
271 s->ip = AUBIO_ARRAY(int , s->fft_size);
272 s->w = AUBIO_ARRAY(smpl_t, s->fft_size);
274 #endif /* using OOURA */
283 void del_aubio_fft(aubio_fft_t * s) {
285 #ifdef HAVE_FFTW3 // using FFTW3
286 pthread_mutex_lock(&aubio_fftw_mutex);
287 fftw_destroy_plan(s->pfw);
288 fftw_destroy_plan(s->pbw);
289 fftw_free(s->specdata);
290 pthread_mutex_unlock(&aubio_fftw_mutex);
292 #elif defined HAVE_ACCELERATE // using ACCELERATE
293 AUBIO_FREE(s->spec.realp);
294 AUBIO_FREE(s->spec.imagp);
295 aubio_vDSP_destroy_fftsetup(s->fftSetup);
297 #elif defined HAVE_INTEL_IPP // using Intel IPP
300 ippFree(s->memBuffer);
301 ippFree(s->complexOut);
308 del_fvec(s->compspec);
314 void aubio_fft_do(aubio_fft_t * s, const fvec_t * input, cvec_t * spectrum) {
315 aubio_fft_do_complex(s, input, s->compspec);
316 aubio_fft_get_spectrum(s->compspec, spectrum);
319 void aubio_fft_rdo(aubio_fft_t * s, const cvec_t * spectrum, fvec_t * output) {
320 aubio_fft_get_realimag(spectrum, s->compspec);
321 aubio_fft_rdo_complex(s, s->compspec, output);
324 void aubio_fft_do_complex(aubio_fft_t * s, const fvec_t * input, fvec_t * compspec) {
326 #ifndef HAVE_MEMCPY_HACKS
327 for (i=0; i < s->winsize; i++) {
328 s->in[i] = input->data[i];
331 memcpy(s->in, input->data, s->winsize * sizeof(smpl_t));
332 #endif /* HAVE_MEMCPY_HACKS */
334 #ifdef HAVE_FFTW3 // using FFTW3
335 fftw_execute(s->pfw);
336 #ifdef HAVE_COMPLEX_H
337 compspec->data[0] = REAL(s->specdata[0]);
338 for (i = 1; i < s->fft_size -1 ; i++) {
339 compspec->data[i] = REAL(s->specdata[i]);
340 compspec->data[compspec->length - i] = IMAG(s->specdata[i]);
342 compspec->data[s->fft_size-1] = REAL(s->specdata[s->fft_size-1]);
343 #else /* HAVE_COMPLEX_H */
344 for (i = 0; i < s->fft_size; i++) {
345 compspec->data[i] = s->specdata[i];
347 #endif /* HAVE_COMPLEX_H */
349 #elif defined HAVE_ACCELERATE // using ACCELERATE
350 // convert real data to even/odd format used in vDSP
351 aubio_vDSP_ctoz((aubio_DSPComplex*)s->in, 2, &s->spec, 1, s->fft_size/2);
353 aubio_vDSP_fft_zrip(s->fftSetup, &s->spec, 1, s->log2fftsize, FFT_FORWARD);
354 // convert from vDSP complex split to [ r0, r1, ..., rN, iN-1, .., i2, i1]
355 compspec->data[0] = s->spec.realp[0];
356 compspec->data[s->fft_size / 2] = s->spec.imagp[0];
357 for (i = 1; i < s->fft_size / 2; i++) {
358 compspec->data[i] = s->spec.realp[i];
359 compspec->data[s->fft_size - i] = s->spec.imagp[i];
362 smpl_t scale = 1./2.;
363 aubio_vDSP_vsmul(compspec->data, 1, &scale, compspec->data, 1, s->fft_size);
365 #elif defined HAVE_INTEL_IPP // using Intel IPP
368 aubio_ippsFFTFwd_RToCCS(s->in, (aubio_IppFloat*)s->complexOut, s->fftSpec, s->memBuffer);
369 // convert complex buffer to [ r0, r1, ..., rN, iN-1, .., i2, i1]
370 compspec->data[0] = s->complexOut[0].re;
371 compspec->data[s->fft_size / 2] = s->complexOut[s->fft_size / 2].re;
372 for (i = 1; i < s->fft_size / 2; i++) {
373 compspec->data[i] = s->complexOut[i].re;
374 compspec->data[s->fft_size - i] = s->complexOut[i].im;
378 aubio_ooura_rdft(s->winsize, 1, s->in, s->ip, s->w);
379 compspec->data[0] = s->in[0];
380 compspec->data[s->winsize / 2] = s->in[1];
381 for (i = 1; i < s->fft_size - 1; i++) {
382 compspec->data[i] = s->in[2 * i];
383 compspec->data[s->winsize - i] = - s->in[2 * i + 1];
385 #endif /* using OOURA */
388 void aubio_fft_rdo_complex(aubio_fft_t * s, const fvec_t * compspec, fvec_t * output) {
391 const smpl_t renorm = 1./(smpl_t)s->winsize;
392 #ifdef HAVE_COMPLEX_H
393 s->specdata[0] = compspec->data[0];
394 for (i=1; i < s->fft_size - 1; i++) {
395 s->specdata[i] = compspec->data[i] +
396 I * compspec->data[compspec->length - i];
398 s->specdata[s->fft_size - 1] = compspec->data[s->fft_size - 1];
400 for (i=0; i < s->fft_size; i++) {
401 s->specdata[i] = compspec->data[i];
404 fftw_execute(s->pbw);
405 for (i = 0; i < output->length; i++) {
406 output->data[i] = s->out[i]*renorm;
409 #elif defined HAVE_ACCELERATE // using ACCELERATE
410 // convert from real imag [ r0, r1, ..., rN, iN-1, .., i2, i1]
411 // to vDSP packed format [ r0, rN, r1, i1, ..., rN-1, iN-1 ]
412 s->out[0] = compspec->data[0];
413 s->out[1] = compspec->data[s->winsize / 2];
414 for (i = 1; i < s->fft_size / 2; i++) {
415 s->out[2 * i] = compspec->data[i];
416 s->out[2 * i + 1] = compspec->data[s->winsize - i];
418 // convert to split complex format used in vDSP
419 aubio_vDSP_ctoz((aubio_DSPComplex*)s->out, 2, &s->spec, 1, s->fft_size/2);
421 aubio_vDSP_fft_zrip(s->fftSetup, &s->spec, 1, s->log2fftsize, FFT_INVERSE);
422 // convert result to real output
423 aubio_vDSP_ztoc(&s->spec, 1, (aubio_DSPComplex*)output->data, 2, s->fft_size/2);
425 smpl_t scale = 1.0 / s->winsize;
426 aubio_vDSP_vsmul(output->data, 1, &scale, output->data, 1, s->fft_size);
428 #elif defined HAVE_INTEL_IPP // using Intel IPP
430 // convert from real imag [ r0, 0, ..., rN, iN-1, .., i2, i1, iN-1] to complex format
431 s->complexOut[0].re = compspec->data[0];
432 s->complexOut[0].im = 0;
433 s->complexOut[s->fft_size / 2].re = compspec->data[s->fft_size / 2];
434 s->complexOut[s->fft_size / 2].im = 0.0;
435 for (i = 1; i < s->fft_size / 2; i++) {
436 s->complexOut[i].re = compspec->data[i];
437 s->complexOut[i].im = compspec->data[s->fft_size - i];
440 aubio_ippsFFTInv_CCSToR((const aubio_IppFloat *)s->complexOut, output->data, s->fftSpec, s->memBuffer);
442 aubio_ippsMulC(output->data, 1.0 / s->winsize, output->data, s->fft_size);
445 smpl_t scale = 2.0 / s->winsize;
446 s->out[0] = compspec->data[0];
447 s->out[1] = compspec->data[s->winsize / 2];
448 for (i = 1; i < s->fft_size - 1; i++) {
449 s->out[2 * i] = compspec->data[i];
450 s->out[2 * i + 1] = - compspec->data[s->winsize - i];
452 aubio_ooura_rdft(s->winsize, -1, s->out, s->ip, s->w);
453 for (i=0; i < s->winsize; i++) {
454 output->data[i] = s->out[i] * scale;
459 void aubio_fft_get_spectrum(const fvec_t * compspec, cvec_t * spectrum) {
460 aubio_fft_get_phas(compspec, spectrum);
461 aubio_fft_get_norm(compspec, spectrum);
464 void aubio_fft_get_realimag(const cvec_t * spectrum, fvec_t * compspec) {
465 aubio_fft_get_imag(spectrum, compspec);
466 aubio_fft_get_real(spectrum, compspec);
469 void aubio_fft_get_phas(const fvec_t * compspec, cvec_t * spectrum) {
471 if (compspec->data[0] < 0) {
472 spectrum->phas[0] = PI;
474 spectrum->phas[0] = 0.;
476 #if defined(HAVE_INTEL_IPP)
477 // convert from real imag [ r0, r1, ..., rN, iN-1, ..., i2, i1, i0]
478 // to [ r0, r1, ..., rN, i0, i1, i2, ..., iN-1]
479 for (i = 1; i < spectrum->length / 2; i++) {
480 ELEM_SWAP(compspec->data[compspec->length - i],
481 compspec->data[spectrum->length + i - 1]);
483 aubio_ippsAtan2(compspec->data + spectrum->length,
484 compspec->data + 1, spectrum->phas + 1, spectrum->length - 1);
485 // revert the imaginary part back again
486 for (i = 1; i < spectrum->length / 2; i++) {
487 ELEM_SWAP(compspec->data[spectrum->length + i - 1],
488 compspec->data[compspec->length - i]);
491 for (i=1; i < spectrum->length - 1; i++) {
492 spectrum->phas[i] = ATAN2(compspec->data[compspec->length-i],
497 // for even length only, make sure last element is 0 or PI
498 if (2 * (compspec->length / 2) == compspec->length) {
500 if (compspec->data[compspec->length/2] < 0) {
501 spectrum->phas[spectrum->length - 1] = PI;
503 spectrum->phas[spectrum->length - 1] = 0.;
507 i = spectrum->length - 1;
508 spectrum->phas[i] = ATAN2(compspec->data[compspec->length-i],
514 void aubio_fft_get_norm(const fvec_t * compspec, cvec_t * spectrum) {
516 spectrum->norm[0] = ABS(compspec->data[0]);
517 for (i=1; i < spectrum->length - 1; i++) {
518 spectrum->norm[i] = SQRT(SQR(compspec->data[i])
519 + SQR(compspec->data[compspec->length - i]) );
522 // for even length, make sure last element is > 0
523 if (2 * (compspec->length / 2) == compspec->length) {
525 spectrum->norm[spectrum->length-1] =
526 ABS(compspec->data[compspec->length/2]);
529 i = spectrum->length - 1;
530 spectrum->norm[i] = SQRT(SQR(compspec->data[i])
531 + SQR(compspec->data[compspec->length - i]) );
536 void aubio_fft_get_imag(const cvec_t * spectrum, fvec_t * compspec) {
538 for (i = 1; i < ( compspec->length + 1 ) / 2 /*- 1 + 1*/; i++) {
539 compspec->data[compspec->length - i] =
540 spectrum->norm[i]*SIN(spectrum->phas[i]);
544 void aubio_fft_get_real(const cvec_t * spectrum, fvec_t * compspec) {
546 for (i = 0; i < compspec->length / 2 + 1; i++) {
548 spectrum->norm[i]*COS(spectrum->phas[i]);