Changeset bfbfafa for src/spectral

Timestamp:

Oct 3, 2017, 10:31:12 PM (7 years ago)

Author:

Paul Brossier <piem@piem.org>

Branches:

feature/autosink, feature/cnn, feature/cnn_org, feature/constantq, feature/crepe, feature/crepe_org, feature/pitchshift, feature/pydocstrings, feature/timestretch, fix/ffmpeg5, master

Children:

2e5c52e

Parents:

7b7a58e (diff), 25db68c (diff)
Note: this is a merge changeset, the changes displayed below correspond to the merge itself.
Use the (diff) links above to see all the changes relative to each parent.

Message:

Merge branch 'master' into dct

File:

• src/spectral/fft.c (modified) (16 diffs)

src/spectral/fft.c

@@ -78 +78 @@
 pthread_mutex_t aubio_fftw_mutex = PTHREAD_MUTEX_INITIALIZER;
 
-#else
-#ifdef HAVE_ACCELERATE        // using ACCELERATE
+#elif defined HAVE_ACCELERATE        // using ACCELERATE
 // https://developer.apple.com/library/mac/#documentation/Accelerate/Reference/vDSPRef/Reference/reference.html
 #include <Accelerate/Accelerate.h>
@@ -113 +112 @@
 #endif /* HAVE_AUBIO_DOUBLE */
 
-#else                         // using OOURA
+#elif defined HAVE_INTEL_IPP // using INTEL IPP
+
+#if !HAVE_AUBIO_DOUBLE
+#define aubio_IppFloat                 Ipp32f
+#define aubio_IppComplex               Ipp32fc
+#define aubio_FFTSpec                  FFTSpec_R_32f
+#define aubio_ippsMalloc_complex       ippsMalloc_32fc
+#define aubio_ippsFFTInit_R            ippsFFTInit_R_32f
+#define aubio_ippsFFTGetSize_R         ippsFFTGetSize_R_32f
+#define aubio_ippsFFTInv_CCSToR        ippsFFTInv_CCSToR_32f
+#define aubio_ippsFFTFwd_RToCCS        ippsFFTFwd_RToCCS_32f
+#define aubio_ippsAtan2                ippsAtan2_32f_A21
+#else /* HAVE_AUBIO_DOUBLE */
+#define aubio_IppFloat                 Ipp64f
+#define aubio_IppComplex               Ipp64fc
+#define aubio_FFTSpec                  FFTSpec_R_64f
+#define aubio_ippsMalloc_complex       ippsMalloc_64fc
+#define aubio_ippsFFTInit_R            ippsFFTInit_R_64f
+#define aubio_ippsFFTGetSize_R         ippsFFTGetSize_R_64f
+#define aubio_ippsFFTInv_CCSToR        ippsFFTInv_CCSToR_64f
+#define aubio_ippsFFTFwd_RToCCS        ippsFFTFwd_RToCCS_64f
+#define aubio_ippsAtan2                ippsAtan2_64f_A50
+#endif
+
+
+#else // using OOURA
 // let's use ooura instead
 extern void aubio_ooura_rdft(int, int, smpl_t *, int *, smpl_t *);
 
-#endif /* HAVE_ACCELERATE */
-#endif /* HAVE_FFTW3 */
+#endif
 
 struct _aubio_fft_t {
   uint_t winsize;
   uint_t fft_size;
+
 #ifdef HAVE_FFTW3             // using FFTW3
   real_t *in, *out;
   fftw_plan pfw, pbw;
-  fft_data_t * specdata;      /* complex spectral data */
-#else
-#ifdef HAVE_ACCELERATE        // using ACCELERATE
+  fft_data_t * specdata; /* complex spectral data */
+
+#elif defined HAVE_ACCELERATE  // using ACCELERATE
   int log2fftsize;
   aubio_FFTSetup fftSetup;
   aubio_DSPSplitComplex spec;
   smpl_t *in, *out;
+
+#elif defined HAVE_INTEL_IPP  // using Intel IPP
+  smpl_t *in, *out;
+  Ipp8u* memSpec;
+  Ipp8u* memInit;
+  Ipp8u* memBuffer;
+  struct aubio_FFTSpec* fftSpec;
+  aubio_IppComplex* complexOut;
 #else                         // using OOURA
   smpl_t *in, *out;
   smpl_t *w;
   int *ip;
-#endif /* HAVE_ACCELERATE */
-#endif /* HAVE_FFTW3 */
+#endif /* using OOURA */
+
   fvec_t * compspec;
 };
@@ -148 +180 @@
     goto beach;
   }
+
 #ifdef HAVE_FFTW3
   uint_t i;
@@ -176 +209 @@
     s->specdata[i] = 0.;
   }
-#else
-#ifdef HAVE_ACCELERATE        // using ACCELERATE
+
+#elif defined HAVE_ACCELERATE  // using ACCELERATE
   s->winsize = winsize;
   s->fft_size = winsize;
   s->compspec = new_fvec(winsize);
-  s->log2fftsize = (uint_t)log2f(s->fft_size);
+  s->log2fftsize = aubio_power_of_two_order(s->fft_size);
   s->in = AUBIO_ARRAY(smpl_t, s->fft_size);
   s->out = AUBIO_ARRAY(smpl_t, s->fft_size);
@@ -187 +220 @@
   s->spec.imagp = AUBIO_ARRAY(smpl_t, s->fft_size/2);
   s->fftSetup = aubio_vDSP_create_fftsetup(s->log2fftsize, FFT_RADIX2);
+
+#elif defined HAVE_INTEL_IPP  // using Intel IPP
+  const IppHintAlgorithm qualityHint = ippAlgHintAccurate; // OR ippAlgHintFast;
+  const int flags = IPP_FFT_NODIV_BY_ANY; // we're scaling manually afterwards
+  int order = aubio_power_of_two_order(winsize);
+  int sizeSpec, sizeInit, sizeBuffer;
+  IppStatus status;
+
+  if (winsize <= 4 || aubio_is_power_of_two(winsize) != 1)
+  {
+    AUBIO_ERR("intel IPP fft: can only create with sizes > 4 and power of two, requested %d,"
+      " try recompiling aubio with --enable-fftw3\n", winsize);
+    goto beach;
+  }
+
+  status = aubio_ippsFFTGetSize_R(order, flags, qualityHint,
+      &sizeSpec, &sizeInit, &sizeBuffer);
+  if (status != ippStsNoErr) {
+    AUBIO_ERR("fft: failed to initialize fft. IPP error: %d\n", status);
+    goto beach;
+  }
+  s->fft_size = s->winsize = winsize;
+  s->compspec = new_fvec(winsize);
+  s->in = AUBIO_ARRAY(smpl_t, s->winsize);
+  s->out = AUBIO_ARRAY(smpl_t, s->winsize);
+  s->memSpec = ippsMalloc_8u(sizeSpec);
+  s->memBuffer = ippsMalloc_8u(sizeBuffer);
+  if (sizeInit > 0 ) {
+    s->memInit = ippsMalloc_8u(sizeInit);
+  }
+  s->complexOut = aubio_ippsMalloc_complex(s->fft_size / 2 + 1);
+  status = aubio_ippsFFTInit_R(
+    &s->fftSpec, order, flags, qualityHint, s->memSpec, s->memInit);
+  if (status != ippStsNoErr) {
+    AUBIO_ERR("fft: failed to initialize. IPP error: %d\n", status);
+    goto beach;
+  }
+
 #else                         // using OOURA
   if (aubio_is_power_of_two(winsize) != 1) {
@@ -201 +272 @@
   s->w     = AUBIO_ARRAY(smpl_t, s->fft_size);
   s->ip[0] = 0;
-#endif /* HAVE_ACCELERATE */
-#endif /* HAVE_FFTW3 */
+#endif /* using OOURA */
+
   return s;
+
 beach:
   AUBIO_FREE(s);
@@ -211 +283 @@
 void del_aubio_fft(aubio_fft_t * s) {
   /* destroy data */
-  del_fvec(s->compspec);
 #ifdef HAVE_FFTW3             // using FFTW3
   pthread_mutex_lock(&aubio_fftw_mutex);
@@ -218 +289 @@
   fftw_free(s->specdata);
   pthread_mutex_unlock(&aubio_fftw_mutex);
-#else /* HAVE_FFTW3 */
-#ifdef HAVE_ACCELERATE        // using ACCELERATE
+
+#elif defined HAVE_ACCELERATE // using ACCELERATE
   AUBIO_FREE(s->spec.realp);
   AUBIO_FREE(s->spec.imagp);
   aubio_vDSP_destroy_fftsetup(s->fftSetup);
+
+#elif defined HAVE_INTEL_IPP  // using Intel IPP
+  ippFree(s->memSpec);
+  ippFree(s->memInit);
+  ippFree(s->memBuffer);
+  ippFree(s->complexOut);
+
 #else                         // using OOURA
   AUBIO_FREE(s->w);
   AUBIO_FREE(s->ip);
-#endif /* HAVE_ACCELERATE */
-#endif /* HAVE_FFTW3 */
+#endif
+
+  del_fvec(s->compspec);
+  AUBIO_FREE(s->in);
   AUBIO_FREE(s->out);
-  AUBIO_FREE(s->in);
   AUBIO_FREE(s);
 }
@@ -252 +331 @@
   memcpy(s->in, input->data, s->winsize * sizeof(smpl_t));
 #endif /* HAVE_MEMCPY_HACKS */
+
 #ifdef HAVE_FFTW3             // using FFTW3
   fftw_execute(s->pfw);
@@ -266 +346 @@
   }
 #endif /* HAVE_COMPLEX_H */
-#else /* HAVE_FFTW3 */
-#ifdef HAVE_ACCELERATE        // using ACCELERATE
+
+#elif defined HAVE_ACCELERATE // using ACCELERATE
   // convert real data to even/odd format used in vDSP
   aubio_vDSP_ctoz((aubio_DSPComplex*)s->in, 2, &s->spec, 1, s->fft_size/2);
@@ -282 +362 @@
   smpl_t scale = 1./2.;
   aubio_vDSP_vsmul(compspec->data, 1, &scale, compspec->data, 1, s->fft_size);
+
+#elif defined HAVE_INTEL_IPP  // using Intel IPP
+
+  // apply fft
+  aubio_ippsFFTFwd_RToCCS(s->in, (aubio_IppFloat*)s->complexOut, s->fftSpec, s->memBuffer);
+  // convert complex buffer to [ r0, r1, ..., rN, iN-1, .., i2, i1]
+  compspec->data[0] = s->complexOut[0].re;
+  compspec->data[s->fft_size / 2] = s->complexOut[s->fft_size / 2].re;
+  for (i = 1; i < s->fft_size / 2; i++) {
+    compspec->data[i] = s->complexOut[i].re;
+    compspec->data[s->fft_size - i] = s->complexOut[i].im;
+  }
+
 #else                         // using OOURA
   aubio_ooura_rdft(s->winsize, 1, s->in, s->ip, s->w);
@@ -290 +383 @@
     compspec->data[s->winsize - i] = - s->in[2 * i + 1];
   }
-#endif /* HAVE_ACCELERATE */
-#endif /* HAVE_FFTW3 */
+#endif /* using OOURA */
 }
 
@@ -314 +406 @@
     output->data[i] = s->out[i]*renorm;
   }
-#else /* HAVE_FFTW3 */
-#ifdef HAVE_ACCELERATE        // using ACCELERATE
+
+#elif defined HAVE_ACCELERATE // using ACCELERATE
   // convert from real imag  [ r0, r1, ..., rN, iN-1, .., i2, i1]
   // to vDSP packed format   [ r0, rN, r1, i1, ..., rN-1, iN-1 ]
@@ -333 +425 @@
   smpl_t scale = 1.0 / s->winsize;
   aubio_vDSP_vsmul(output->data, 1, &scale, output->data, 1, s->fft_size);
+
+#elif defined HAVE_INTEL_IPP  // using Intel IPP
+
+  // convert from real imag  [ r0, 0, ..., rN, iN-1, .., i2, i1, iN-1] to complex format
+  s->complexOut[0].re = compspec->data[0];
+  s->complexOut[0].im = 0;
+  s->complexOut[s->fft_size / 2].re = compspec->data[s->fft_size / 2];
+  s->complexOut[s->fft_size / 2].im = 0.0;
+  for (i = 1; i < s->fft_size / 2; i++) {
+    s->complexOut[i].re = compspec->data[i];
+    s->complexOut[i].im = compspec->data[s->fft_size - i];
+  }
+  // apply fft
+  aubio_ippsFFTInv_CCSToR((const aubio_IppFloat *)s->complexOut, output->data, s->fftSpec, s->memBuffer);
+  // apply scaling
+  aubio_ippsMulC(output->data, 1.0 / s->winsize, output->data, s->fft_size);
+
 #else                         // using OOURA
   smpl_t scale = 2.0 / s->winsize;
@@ -345 +454 @@
     output->data[i] = s->out[i] * scale;
   }
-#endif /* HAVE_ACCELERATE */
-#endif /* HAVE_FFTW3 */
+#endif
 }
 
@@ -366 +474 @@
     spectrum->phas[0] = 0.;
   }
+#if defined(HAVE_INTEL_IPP)
+  // convert from real imag  [ r0, r1, ..., rN, iN-1, ..., i2, i1, i0]
+  //                     to  [ r0, r1, ..., rN, i0, i1, i2, ..., iN-1]
+  for (i = 1; i < spectrum->length / 2; i++) {
+    ELEM_SWAP(compspec->data[compspec->length - i],
+        compspec->data[spectrum->length + i - 1]);
+  }
+  aubio_ippsAtan2(compspec->data + spectrum->length,
+      compspec->data + 1, spectrum->phas + 1, spectrum->length - 1);
+  // revert the imaginary part back again
+  for (i = 1; i < spectrum->length / 2; i++) {
+    ELEM_SWAP(compspec->data[spectrum->length + i - 1],
+        compspec->data[compspec->length - i]);
+  }
+#else
   for (i=1; i < spectrum->length - 1; i++) {
     spectrum->phas[i] = ATAN2(compspec->data[compspec->length-i],
         compspec->data[i]);
   }
+#endif
   if (compspec->data[compspec->length/2] < 0) {
     spectrum->phas[spectrum->length - 1] = PI;