changeset 434:e979a9c4ffb6 vampipe

Switch from Cross FFT with option of FFTW build, to KissFFT only (code bundled). This is much faster than the default build and simpler than managing two options.

--- a/Makefile.in	Wed May 18 12:21:17 2016 +0100
+++ b/Makefile.in	Tue Aug 16 16:04:09 2016 +0100
@@ -390,7 +390,7 @@
 src/vamp-sdk/PluginAdapter.o: vamp-sdk/Plugin.h vamp-sdk/PluginBase.h
 src/vamp-sdk/PluginAdapter.o: vamp-sdk/plugguard.h vamp-sdk/RealTime.h
 src/vamp-sdk/RealTime.o: ./vamp-sdk/RealTime.h vamp-sdk/plugguard.h
-src/vamp-sdk/FFT.o: src/vamp-sdk/FFT.cpp vamp-sdk/FFT.h src/vamp-sdk/FFTimpl.cpp
+src/vamp-sdk/FFT.o: src/vamp-sdk/FFT.cpp vamp-sdk/FFT.h 
 src/vamp-hostsdk/PluginBufferingAdapter.o: ./vamp-hostsdk/PluginBufferingAdapter.h
 src/vamp-hostsdk/PluginBufferingAdapter.o: ./vamp-hostsdk/PluginWrapper.h
 src/vamp-hostsdk/PluginBufferingAdapter.o: ./vamp-hostsdk/Plugin.h
@@ -416,7 +416,6 @@
 src/vamp-hostsdk/PluginInputDomainAdapter.o: vamp-sdk/plugguard.h
 src/vamp-hostsdk/PluginInputDomainAdapter.o: vamp-sdk/RealTime.h
 src/vamp-hostsdk/PluginInputDomainAdapter.o: vamp-sdk/FFT.h
-src/vamp-hostsdk/PluginInputDomainAdapter.o: src/vamp-sdk/FFTimpl.cpp
 src/vamp-hostsdk/PluginLoader.o: ./vamp-hostsdk/PluginHostAdapter.h
 src/vamp-hostsdk/PluginLoader.o: vamp/vamp.h vamp-sdk/Plugin.h
 src/vamp-hostsdk/PluginLoader.o: vamp-sdk/PluginBase.h

--- a/src/vamp-hostsdk/PluginInputDomainAdapter.cpp	Wed May 18 12:21:17 2016 +0100
+++ b/src/vamp-hostsdk/PluginInputDomainAdapter.cpp	Tue Aug 16 16:04:09 2016 +0100
@@ -43,37 +43,29 @@
 
 #include "Window.h"
 
+#include <stdlib.h>
+#include <stdio.h>
+#include <math.h>
+#include <string.h>
+#include <limits.h>
 
-/**
- * If you want to compile using FFTW instead of the built-in FFT
- * implementation for the PluginInputDomainAdapter, define HAVE_FFTW3
- * in the Makefile.
- *
- * Be aware that FFTW is licensed under the GPL -- unlike this SDK,
- * which is provided under a more liberal BSD license in order to
- * permit use in closed source applications.  The use of FFTW would
- * mean that your code would need to be licensed under the GPL as
- * well.  Do not define this symbol unless you understand and accept
- * the implications of this.
- *
- * Parties such as Linux distribution packagers who redistribute this
- * SDK for use in other programs should _not_ define this symbol, as
- * it would change the effective licensing terms under which the SDK
- * was available to third party developers.
- *
- * The default is not to use FFTW, and to use the built-in FFT instead.
- * 
- * Note: The FFTW code uses FFTW_MEASURE, and so will perform badly on
- * its first invocation unless the host has saved and restored FFTW
- * wisdom (see the FFTW documentation).
- */
-#ifdef HAVE_FFTW3
-#include <fftw3.h>
-#pragma message("*** NOTE: Compiling with FFTW3 support will result in a GPL binary")
-#else
-#include "../vamp-sdk/FFTimpl.cpp"
-#endif
+// Override C linkage for KissFFT headers. So long as we have already
+// included all of the other (system etc) headers KissFFT depends on,
+// this should work out OK
+#undef __cplusplus
 
+namespace KissSingle {
+#undef KISS_FFT_H
+#undef KISS_FTR_H
+#undef KISS_FFT__GUTS_H
+#undef FIXED_POINT
+#undef USE_SIMD
+#undef kiss_fft_scalar
+#define kiss_fft_scalar float
+inline void free(void *ptr) { ::free(ptr); }
+#include "../vamp-sdk/ext/kiss_fft.c"
+#include "../vamp-sdk/ext/kiss_fftr.c"
+}
 
 _VAMP_SDK_HOSTSPACE_BEGIN(PluginInputDomainAdapter.cpp)
 
@@ -110,30 +102,24 @@
     int m_stepSize;
     int m_blockSize;
     float **m_freqbuf;
-
-    double *m_ri;
+    float *m_ri;
 
     WindowType m_windowType;
-    Window<double> *m_window;
+    Window<float> *m_window;
 
     ProcessTimestampMethod m_method;
     int m_processCount;
     float **m_shiftBuffers;
 
-#ifdef HAVE_FFTW3
-    fftw_plan m_plan;
-    fftw_complex *m_cbuf;
-#else
-    double *m_ro;
-    double *m_io;
-#endif
+    KissSingle::kiss_fftr_cfg m_cfg;
+    KissSingle::kiss_fft_cpx *m_cbuf;
 
     FeatureSet processShiftingTimestamp(const float *const *inputBuffers, RealTime timestamp);
     FeatureSet processShiftingData(const float *const *inputBuffers, RealTime timestamp);
 
     size_t makeBlockSizeAcceptable(size_t) const;
     
-    Window<double>::WindowType convertType(WindowType t) const;
+    Window<float>::WindowType convertType(WindowType t) const;
 };
 
 PluginInputDomainAdapter::PluginInputDomainAdapter(Plugin *plugin) :
@@ -227,13 +213,8 @@
     m_method(ShiftTimestamp),
     m_processCount(0),
     m_shiftBuffers(0),
-#ifdef HAVE_FFTW3
-    m_plan(0),
+    m_cfg(0),
     m_cbuf(0)
-#else
-    m_ro(0),
-    m_io(0)
-#endif
 {
 }
 
@@ -253,19 +234,11 @@
             delete[] m_freqbuf[c];
         }
         delete[] m_freqbuf;
-#ifdef HAVE_FFTW3
-        if (m_plan) {
-            fftw_destroy_plan(m_plan);
-            fftw_free(m_ri);
-            fftw_free(m_cbuf);
-            m_plan = 0;
+        if (m_cfg) {
+            KissSingle::kiss_fftr_free(m_cfg);
+            delete[] m_cbuf;
+            m_cfg = 0;
         }
-#else
-        delete[] m_ri;
-        delete[] m_ro;
-        delete[] m_io;
-#endif
-
         delete m_window;
     }
 }
@@ -292,30 +265,21 @@
         return false;
     }                
         
-#ifndef HAVE_FFTW3
-    if (blockSize & (blockSize-1)) {
-        std::cerr << "ERROR: PluginInputDomainAdapter::initialise: non-power-of-two\nblocksize " << blockSize << " not supported" << std::endl;
+    if (blockSize % 2) {
+        std::cerr << "ERROR: PluginInputDomainAdapter::initialise: odd blocksize " << blockSize << " not supported" << std::endl;
         return false;
     }
-#endif
 
     if (m_channels > 0) {
         for (int c = 0; c < m_channels; ++c) {
             delete[] m_freqbuf[c];
         }
         delete[] m_freqbuf;
-#ifdef HAVE_FFTW3
-        if (m_plan) {
-            fftw_destroy_plan(m_plan);
-            fftw_free(m_ri);
-            fftw_free(m_cbuf);
-            m_plan = 0;
+        if (m_cfg) {
+            KissSingle::kiss_fftr_free(m_cfg);
+            delete[] m_cbuf;
+            m_cfg = 0;
         }
-#else
-        delete[] m_ri;
-        delete[] m_ro;
-        delete[] m_io;
-#endif
         delete m_window;
     }
 
@@ -328,17 +292,10 @@
         m_freqbuf[c] = new float[m_blockSize + 2];
     }
 
-    m_window = new Window<double>(convertType(m_windowType), m_blockSize);
+    m_window = new Window<float>(convertType(m_windowType), m_blockSize);
 
-#ifdef HAVE_FFTW3
-    m_ri = (double *)fftw_malloc(blockSize * sizeof(double));
-    m_cbuf = (fftw_complex *)fftw_malloc((blockSize/2 + 1) * sizeof(fftw_complex));
-    m_plan = fftw_plan_dft_r2c_1d(int(blockSize), m_ri, m_cbuf, FFTW_MEASURE);
-#else
-    m_ri = new double[m_blockSize];
-    m_ro = new double[m_blockSize];
-    m_io = new double[m_blockSize];
-#endif
+    m_cfg = KissSingle::kiss_fftr_alloc(blockSize, false, 0, 0);
+    m_cbuf = new KissSingle::kiss_fft_cpx[blockSize/2+1];
 
     m_processCount = 0;
 
@@ -388,36 +345,12 @@
         std::cerr << "WARNING: PluginInputDomainAdapter::initialise: blocksize < 2 not" << std::endl
                   << "supported, increasing from " << blockSize << " to 2" << std::endl;
         blockSize = 2;
+
+    } else if (blockSize % 2) {
         
-    } else if (blockSize & (blockSize-1)) {
-            
-#ifdef HAVE_FFTW3
-        // not an issue with FFTW
-#else
-
-        // not a power of two, can't handle that with our built-in FFT
-        // implementation
-
-        size_t nearest = blockSize;
-        size_t power = 0;
-        while (nearest > 1) {
-            nearest >>= 1;
-            ++power;
-        }
-        nearest = 1;
-        while (power) {
-            nearest <<= 1;
-            --power;
-        }
-        
-        if (blockSize - nearest > (nearest*2) - blockSize) {
-            nearest = nearest*2;
-        }
-        
-        std::cerr << "WARNING: PluginInputDomainAdapter::initialise: non-power-of-two\nblocksize " << blockSize << " not supported, using blocksize " << nearest << " instead" << std::endl;
-        blockSize = nearest;
-
-#endif
+        std::cerr << "WARNING: PluginInputDomainAdapter::initialise: odd blocksize not" << std::endl
+                  << "supported, increasing from " << blockSize << " to " << (blockSize+1) << std::endl;
+        blockSize = blockSize+1;
     }
 
     return blockSize;
@@ -455,7 +388,7 @@
     m_windowType = t;
     if (m_window) {
         delete m_window;
-        m_window = new Window<double>(convertType(m_windowType), m_blockSize);
+        m_window = new Window<float>(convertType(m_windowType), m_blockSize);
     }
 }
 
@@ -465,26 +398,26 @@
     return m_windowType;
 }
 
-Window<double>::WindowType
+Window<float>::WindowType
 PluginInputDomainAdapter::Impl::convertType(WindowType t) const
 {
     switch (t) {
     case RectangularWindow:
-        return Window<double>::RectangularWindow;
+        return Window<float>::RectangularWindow;
     case BartlettWindow:
-        return Window<double>::BartlettWindow;
+        return Window<float>::BartlettWindow;
     case HammingWindow:
-        return Window<double>::HammingWindow;
+        return Window<float>::HammingWindow;
     case HanningWindow:
-        return Window<double>::HanningWindow;
+        return Window<float>::HanningWindow;
     case BlackmanWindow:
-        return Window<double>::BlackmanWindow;
+        return Window<float>::BlackmanWindow;
     case NuttallWindow:
-        return Window<double>::NuttallWindow;
+        return Window<float>::NuttallWindow;
     case BlackmanHarrisWindow:
-        return Window<double>::BlackmanHarrisWindow;
+        return Window<float>::BlackmanHarrisWindow;
     default:
-	return Window<double>::HanningWindow;
+	return Window<float>::HanningWindow;
     }
 }
 
@@ -529,26 +462,17 @@
 
         for (int i = 0; i < m_blockSize/2; ++i) {
             // FFT shift
-            double value = m_ri[i];
+            float value = m_ri[i];
             m_ri[i] = m_ri[i + m_blockSize/2];
             m_ri[i + m_blockSize/2] = value;
         }
 
-#ifdef HAVE_FFTW3
-        fftw_execute(m_plan);
-
+        KissSingle::kiss_fftr(m_cfg, m_ri, m_cbuf);
+        
         for (int i = 0; i <= m_blockSize/2; ++i) {
-            m_freqbuf[c][i * 2] = float(m_cbuf[i][0]);
-            m_freqbuf[c][i * 2 + 1] = float(m_cbuf[i][1]);
+            m_freqbuf[c][i * 2] = m_cbuf[i].r;
+            m_freqbuf[c][i * 2 + 1] = m_cbuf[i].i;
         }
-#else
-        fft(m_blockSize, false, m_ri, 0, m_ro, m_io);
-
-        for (int i = 0; i <= m_blockSize/2; ++i) {
-            m_freqbuf[c][i * 2] = float(m_ro[i]);
-            m_freqbuf[c][i * 2 + 1] = float(m_io[i]);
-        }
-#endif
     }
 
     return m_plugin->process(m_freqbuf, timestamp);
@@ -587,26 +511,17 @@
 
         for (int i = 0; i < m_blockSize/2; ++i) {
             // FFT shift
-            double value = m_ri[i];
+            float value = m_ri[i];
             m_ri[i] = m_ri[i + m_blockSize/2];
             m_ri[i + m_blockSize/2] = value;
         }
 
-#ifdef HAVE_FFTW3
-        fftw_execute(m_plan);
-
+        KissSingle::kiss_fftr(m_cfg, m_ri, m_cbuf);
+        
         for (int i = 0; i <= m_blockSize/2; ++i) {
-            m_freqbuf[c][i * 2] = float(m_cbuf[i][0]);
-            m_freqbuf[c][i * 2 + 1] = float(m_cbuf[i][1]);
+            m_freqbuf[c][i * 2] = m_cbuf[i].r;
+            m_freqbuf[c][i * 2 + 1] = m_cbuf[i].i;
         }
-#else
-        fft(m_blockSize, false, m_ri, 0, m_ro, m_io);
-
-        for (int i = 0; i <= m_blockSize/2; ++i) {
-            m_freqbuf[c][i * 2] = float(m_ro[i]);
-            m_freqbuf[c][i * 2 + 1] = float(m_io[i]);
-        }
-#endif
     }
 
     ++m_processCount;
@@ -614,10 +529,6 @@
     return m_plugin->process(m_freqbuf, timestamp);
 }
 
-#ifndef HAVE_FFTW3
-
-#endif
-
 }
         
 }

--- a/src/vamp-sdk/FFT.cpp	Wed May 18 12:21:17 2016 +0100
+++ b/src/vamp-sdk/FFT.cpp	Tue Aug 16 16:04:09 2016 +0100
@@ -36,37 +36,173 @@
 
 #include <vamp-sdk/FFT.h>
 
-#include <cmath>
+#include <stdlib.h>
+#include <stdio.h>
+#include <math.h>
+#include <string.h>
 
 #if ( VAMP_SDK_MAJOR_VERSION != 2 || VAMP_SDK_MINOR_VERSION != 7 )
 #error Unexpected version of Vamp SDK header included
 #endif
 
-#ifdef _MSC_VER
-#include <stdlib.h>
-#include <malloc.h>
-#endif
+_VAMP_SDK_PLUGSPACE_BEGIN(FFT.cpp)
 
-_VAMP_SDK_PLUGSPACE_BEGIN(FFT.cpp)
+// Override C linkage for KissFFT headers. So long as we have already
+// included all of the other (system etc) headers KissFFT depends on,
+// this should work out OK
+#undef __cplusplus
+
+namespace KissSingle {
+#undef KISS_FFT_H
+#undef KISS_FTR_H
+#undef KISS_FFT__GUTS_H
+#undef FIXED_POINT
+#undef USE_SIMD
+#undef kiss_fft_scalar
+#define kiss_fft_scalar float
+inline void free(void *ptr) { ::free(ptr); }
+#include "ext/kiss_fft.c"
+#include "ext/kiss_fftr.c"
+}
+
+namespace KissDouble {
+#undef KISS_FFT_H
+#undef KISS_FTR_H
+#undef KISS_FFT__GUTS_H
+#undef FIXED_POINT
+#undef USE_SIMD
+#undef kiss_fft_scalar
+#define kiss_fft_scalar double
+inline void free(void *ptr) { ::free(ptr); }
+#include "ext/kiss_fft.c"
+#include "ext/kiss_fftr.c"
+}
 
 namespace Vamp {
 
-#include "FFTimpl.cpp"
-
 void
-FFT::forward(unsigned int n,
+FFT::forward(unsigned int un,
 	     const double *ri, const double *ii,
 	     double *ro, double *io)
 {
-    fft(n, false, ri, ii, ro, io);
+    int n(un);
+    KissDouble::kiss_fft_cfg c = KissDouble::kiss_fft_alloc(n, false, 0, 0);
+    KissDouble::kiss_fft_cpx *in = new KissDouble::kiss_fft_cpx[n];
+    KissDouble::kiss_fft_cpx *out = new KissDouble::kiss_fft_cpx[n];
+    for (int i = 0; i < n; ++i) {
+        in[i].r = ri[i];
+        in[i].i = 0;
+    }
+    if (ii) {
+        for (int i = 0; i < n; ++i) {
+            in[i].i = ii[i];
+        }
+    }
+    kiss_fft(c, in, out);
+    for (int i = 0; i < n; ++i) {
+        ro[i] = out[i].r;
+        io[i] = out[i].i;
+    }
+    KissDouble::kiss_fft_free(c);
+    delete[] in;
+    delete[] out;
 }
 
 void
-FFT::inverse(unsigned int n,
+FFT::inverse(unsigned int un,
 	     const double *ri, const double *ii,
 	     double *ro, double *io)
 {
-    fft(n, true, ri, ii, ro, io);
+    int n(un);
+    KissDouble::kiss_fft_cfg c = KissDouble::kiss_fft_alloc(n, true, 0, 0);
+    KissDouble::kiss_fft_cpx *in = new KissDouble::kiss_fft_cpx[n];
+    KissDouble::kiss_fft_cpx *out = new KissDouble::kiss_fft_cpx[n];
+    for (int i = 0; i < n; ++i) {
+        in[i].r = ri[i];
+        in[i].i = 0;
+    }
+    if (ii) {
+        for (int i = 0; i < n; ++i) {
+            in[i].i = ii[i];
+        }
+    }
+    kiss_fft(c, in, out);
+    double scale = 1.0 / double(n);
+    for (int i = 0; i < n; ++i) {
+        ro[i] = out[i].r * scale;
+        io[i] = out[i].i * scale;
+    }
+    KissDouble::kiss_fft_free(c);
+    delete[] in;
+    delete[] out;
+    KissDouble::kiss_fft_free(c);
+}
+
+class FFTReal::D
+{
+public:
+    
+    D(int n) :
+        m_n(n),
+        m_cf(KissSingle::kiss_fftr_alloc(n, false, 0, 0)),
+        m_ci(KissSingle::kiss_fftr_alloc(n, true, 0, 0)),
+        m_freq(new KissSingle::kiss_fft_cpx[n/2+1]) { }
+
+    ~D() {
+        KissSingle::kiss_fftr_free(m_cf);
+        KissSingle::kiss_fftr_free(m_ci);
+        delete[] m_freq;
+    }
+
+    void forward(const float *ri, float *co) {
+        KissSingle::kiss_fftr(m_cf, ri, m_freq);
+        int hs = m_n/2 + 1;
+        for (int i = 0; i < hs; ++i) {
+            co[i*2] = m_freq[i].r;
+            co[i*2+1] = m_freq[i].i;
+        }
+    }
+
+    void inverse(const float *ci, float *ro) {
+        int hs = m_n/2 + 1;
+        for (int i = 0; i < hs; ++i) {
+            m_freq[i].r = ci[i*2];
+            m_freq[i].i = ci[i*2+1];
+        }
+        KissSingle::kiss_fftri(m_ci, m_freq, ro);
+        double scale = 1.0 / double(m_n);
+        for (int i = 0; i < m_n; ++i) {
+            ro[i] *= scale;
+        }
+    }
+    
+private:
+    int m_n;
+    KissSingle::kiss_fftr_cfg m_cf;
+    KissSingle::kiss_fftr_cfg m_ci;
+    KissSingle::kiss_fft_cpx *m_freq;
+};
+
+FFTReal::FFTReal(unsigned int n) :
+    m_d(new D(n))
+{
+}
+
+FFTReal::~FFTReal()
+{
+    delete m_d;
+}
+
+void
+FFTReal::forward(const float *ri, float *co)
+{
+    m_d->forward(ri, co);
+}
+
+void
+FFTReal::inverse(const float *ci, float *ro)
+{
+    m_d->inverse(ci, ro);
 }
 
 }

--- a/src/vamp-sdk/FFTimpl.cpp	Wed May 18 12:21:17 2016 +0100
+++ /dev/null	Thu Jan 01 00:00:00 1970 +0000
@@ -1,116 +0,0 @@
-
-/* Public domain FFT implementation from Don Cross. */
-
-static void
-fft(unsigned int n, bool inverse,
-    const double *ri, const double *ii,
-    double *ro, double *io)
-{
-    if (!ri || !ro || !io) return;
-
-    unsigned int bits;
-    unsigned int i, j, k, m;
-    unsigned int blockSize, blockEnd;
-
-    double tr, ti;
-
-    if (n < 2) return;
-    if (n & (n-1)) return;
-
-    double angle = 2.0 * M_PI;
-    if (inverse) angle = -angle;
-
-    for (i = 0; ; ++i) {
-	if (n & (1 << i)) {
-	    bits = i;
-	    break;
-	}
-    }
-
-#ifdef _MSC_VER
-    int *table = (int *)_malloca(n * sizeof(int));
-#else
-    int table[n];
-#endif
-
-    for (i = 0; i < n; ++i) {
-        m = i;
-        for (j = k = 0; j < bits; ++j) {
-            k = (k << 1) | (m & 1);
-            m >>= 1;
-        }
-        table[i] = k;
-    }
-
-    if (ii) {
-	for (i = 0; i < n; ++i) {
-	    ro[table[i]] = ri[i];
-	    io[table[i]] = ii[i];
-	}
-    } else {
-	for (i = 0; i < n; ++i) {
-	    ro[table[i]] = ri[i];
-	    io[table[i]] = 0.0;
-	}
-    }
-
-    blockEnd = 1;
-
-    for (blockSize = 2; blockSize <= n; blockSize <<= 1) {
-
-	double delta = angle / (double)blockSize;
-	double sm2 = -sin(-2 * delta);
-	double sm1 = -sin(-delta);
-	double cm2 = cos(-2 * delta);
-	double cm1 = cos(-delta);
-	double w = 2 * cm1;
-	double ar[3], ai[3];
-
-	for (i = 0; i < n; i += blockSize) {
-
-	    ar[2] = cm2;
-	    ar[1] = cm1;
-
-	    ai[2] = sm2;
-	    ai[1] = sm1;
-
-	    for (j = i, m = 0; m < blockEnd; j++, m++) {
-
-		ar[0] = w * ar[1] - ar[2];
-		ar[2] = ar[1];
-		ar[1] = ar[0];
-
-		ai[0] = w * ai[1] - ai[2];
-		ai[2] = ai[1];
-		ai[1] = ai[0];
-
-		k = j + blockEnd;
-		tr = ar[0] * ro[k] - ai[0] * io[k];
-		ti = ar[0] * io[k] + ai[0] * ro[k];
-
-		ro[k] = ro[j] - tr;
-		io[k] = io[j] - ti;
-
-		ro[j] += tr;
-		io[j] += ti;
-	    }
-	}
-
-	blockEnd = blockSize;
-    }
-
-    if (inverse) {
-
-	double denom = (double)n;
-
-	for (i = 0; i < n; i++) {
-	    ro[i] /= denom;
-	    io[i] /= denom;
-	}
-    }
-
-#ifdef _MSC_VER
-    _freea(table);
-#endif
-}
-

--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/src/vamp-sdk/ext/_kiss_fft_guts.h	Tue Aug 16 16:04:09 2016 +0100
@@ -0,0 +1,168 @@
+#ifndef KISS_FFT__GUTS_H
+#define KISS_FFT__GUTS_H
+/*
+Copyright (c) 2003-2010, Mark Borgerding
+
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
+    * Neither the author nor the names of any contributors may be used to endorse or promote products derived from this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+/* kiss_fft.h
+   defines kiss_fft_scalar as either short or a float type
+   and defines
+   typedef struct { kiss_fft_scalar r; kiss_fft_scalar i; }kiss_fft_cpx; */
+#include "kiss_fft.h"
+#include <limits.h>
+
+#define MAXFACTORS 32
+/* e.g. an fft of length 128 has 4 factors 
+ as far as kissfft is concerned
+ 4*4*4*2
+ */
+
+struct kiss_fft_state{
+    int nfft;
+    int inverse;
+    int factors[2*MAXFACTORS];
+    kiss_fft_cpx twiddles[1];
+};
+
+/*
+  Explanation of macros dealing with complex math:
+
+   C_MUL(m,a,b)         : m = a*b
+   C_FIXDIV( c , div )  : if a fixed point impl., c /= div. noop otherwise
+   C_SUB( res, a,b)     : res = a - b
+   C_SUBFROM( res , a)  : res -= a
+   C_ADDTO( res , a)    : res += a
+ * */
+#ifdef FIXED_POINT
+#if (FIXED_POINT==32)
+# define FRACBITS 31
+# define SAMPPROD int64_t
+#define SAMP_MAX 2147483647
+#else
+# define FRACBITS 15
+# define SAMPPROD int32_t 
+#define SAMP_MAX 32767
+#endif
+
+#define SAMP_MIN -SAMP_MAX
+
+#if defined(CHECK_OVERFLOW)
+#  define CHECK_OVERFLOW_OP(a,op,b)  \
+	if ( (SAMPPROD)(a) op (SAMPPROD)(b) > SAMP_MAX || (SAMPPROD)(a) op (SAMPPROD)(b) < SAMP_MIN ) { \
+		fprintf(stderr,"WARNING:overflow @ " __FILE__ "(%d): (%d " #op" %d) = %ld\n",__LINE__,(a),(b),(SAMPPROD)(a) op (SAMPPROD)(b) );  }
+#endif
+
+
+#   define smul(a,b) ( (SAMPPROD)(a)*(b) )
+#   define sround( x )  (kiss_fft_scalar)( ( (x) + (1<<(FRACBITS-1)) ) >> FRACBITS )
+
+#   define S_MUL(a,b) sround( smul(a,b) )
+
+#   define C_MUL(m,a,b) \
+      do{ (m).r = sround( smul((a).r,(b).r) - smul((a).i,(b).i) ); \
+          (m).i = sround( smul((a).r,(b).i) + smul((a).i,(b).r) ); }while(0)
+
+#   define DIVSCALAR(x,k) \
+	(x) = sround( smul(  x, SAMP_MAX/k ) )
+
+#   define C_FIXDIV(c,div) \
+	do {    DIVSCALAR( (c).r , div);  \
+		DIVSCALAR( (c).i  , div); }while (0)
+
+#   define C_MULBYSCALAR( c, s ) \
+    do{ (c).r =  sround( smul( (c).r , s ) ) ;\
+        (c).i =  sround( smul( (c).i , s ) ) ; }while(0)
+
+#else  /* not FIXED_POINT*/
+
+#   define S_MUL(a,b) ( (a)*(b) )
+#define C_MUL(m,a,b) \
+    do{ (m).r = (a).r*(b).r - (a).i*(b).i;\
+        (m).i = (a).r*(b).i + (a).i*(b).r; }while(0)
+#   define C_FIXDIV(c,div) /* NOOP */
+#   define C_MULBYSCALAR( c, s ) \
+    do{ (c).r *= (s);\
+        (c).i *= (s); }while(0)
+#endif
+
+#ifndef CHECK_OVERFLOW_OP
+#  define CHECK_OVERFLOW_OP(a,op,b) /* noop */
+#endif
+
+#define  C_ADD( res, a,b)\
+    do { \
+	    CHECK_OVERFLOW_OP((a).r,+,(b).r)\
+	    CHECK_OVERFLOW_OP((a).i,+,(b).i)\
+	    (res).r=(a).r+(b).r;  (res).i=(a).i+(b).i; \
+    }while(0)
+#define  C_SUB( res, a,b)\
+    do { \
+	    CHECK_OVERFLOW_OP((a).r,-,(b).r)\
+	    CHECK_OVERFLOW_OP((a).i,-,(b).i)\
+	    (res).r=(a).r-(b).r;  (res).i=(a).i-(b).i; \
+    }while(0)
+#define C_ADDTO( res , a)\
+    do { \
+	    CHECK_OVERFLOW_OP((res).r,+,(a).r)\
+	    CHECK_OVERFLOW_OP((res).i,+,(a).i)\
+	    (res).r += (a).r;  (res).i += (a).i;\
+    }while(0)
+
+#define C_SUBFROM( res , a)\
+    do {\
+	    CHECK_OVERFLOW_OP((res).r,-,(a).r)\
+	    CHECK_OVERFLOW_OP((res).i,-,(a).i)\
+	    (res).r -= (a).r;  (res).i -= (a).i; \
+    }while(0)
+
+
+#ifdef FIXED_POINT
+#  define KISS_FFT_COS(phase)  floor(.5+SAMP_MAX * cos (phase))
+#  define KISS_FFT_SIN(phase)  floor(.5+SAMP_MAX * sin (phase))
+#  define HALF_OF(x) ((x)>>1)
+#elif defined(USE_SIMD)
+#  define KISS_FFT_COS(phase) _mm_set1_ps( cos(phase) )
+#  define KISS_FFT_SIN(phase) _mm_set1_ps( sin(phase) )
+#  define HALF_OF(x) ((x)*_mm_set1_ps(.5))
+#else
+#  define KISS_FFT_COS(phase) (kiss_fft_scalar) cos(phase)
+#  define KISS_FFT_SIN(phase) (kiss_fft_scalar) sin(phase)
+#  define HALF_OF(x) ((x)*.5)
+#endif
+
+#define  kf_cexp(x,phase) \
+	do{ \
+		(x)->r = KISS_FFT_COS(phase);\
+		(x)->i = KISS_FFT_SIN(phase);\
+	}while(0)
+
+
+/* a debugging function */
+#define pcpx(c)\
+    fprintf(stderr,"%g + %gi\n",(double)((c)->r),(double)((c)->i) )
+
+
+#ifdef KISS_FFT_USE_ALLOCA
+// define this to allow use of alloca instead of malloc for temporary buffers
+// Temporary buffers are used in two case: 
+// 1. FFT sizes that have "bad" factors. i.e. not 2,3 and 5
+// 2. "in-place" FFTs.  Notice the quotes, since kissfft does not really do an in-place transform.
+#include <alloca.h>
+#define  KISS_FFT_TMP_ALLOC(nbytes) alloca(nbytes)
+#define  KISS_FFT_TMP_FREE(ptr) 
+#else
+#define  KISS_FFT_TMP_ALLOC(nbytes) KISS_FFT_MALLOC(nbytes)
+#define  KISS_FFT_TMP_FREE(ptr) KISS_FFT_FREE(ptr)
+#endif
+
+#endif

--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/src/vamp-sdk/ext/kiss_fft.c	Tue Aug 16 16:04:09 2016 +0100
@@ -0,0 +1,408 @@
+/*
+Copyright (c) 2003-2010, Mark Borgerding
+
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
+    * Neither the author nor the names of any contributors may be used to endorse or promote products derived from this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+
+#include "_kiss_fft_guts.h"
+/* The guts header contains all the multiplication and addition macros that are defined for
+ fixed or floating point complex numbers.  It also delares the kf_ internal functions.
+ */
+
+static void kf_bfly2(
+        kiss_fft_cpx * Fout,
+        const size_t fstride,
+        const kiss_fft_cfg st,
+        int m
+        )
+{
+    kiss_fft_cpx * Fout2;
+    kiss_fft_cpx * tw1 = st->twiddles;
+    kiss_fft_cpx t;
+    Fout2 = Fout + m;
+    do{
+        C_FIXDIV(*Fout,2); C_FIXDIV(*Fout2,2);
+
+        C_MUL (t,  *Fout2 , *tw1);
+        tw1 += fstride;
+        C_SUB( *Fout2 ,  *Fout , t );
+        C_ADDTO( *Fout ,  t );
+        ++Fout2;
+        ++Fout;
+    }while (--m);
+}
+
+static void kf_bfly4(
+        kiss_fft_cpx * Fout,
+        const size_t fstride,
+        const kiss_fft_cfg st,
+        const size_t m
+        )
+{
+    kiss_fft_cpx *tw1,*tw2,*tw3;
+    kiss_fft_cpx scratch[6];
+    size_t k=m;
+    const size_t m2=2*m;
+    const size_t m3=3*m;
+
+
+    tw3 = tw2 = tw1 = st->twiddles;
+
+    do {
+        C_FIXDIV(*Fout,4); C_FIXDIV(Fout[m],4); C_FIXDIV(Fout[m2],4); C_FIXDIV(Fout[m3],4);
+
+        C_MUL(scratch[0],Fout[m] , *tw1 );
+        C_MUL(scratch[1],Fout[m2] , *tw2 );
+        C_MUL(scratch[2],Fout[m3] , *tw3 );
+
+        C_SUB( scratch[5] , *Fout, scratch[1] );
+        C_ADDTO(*Fout, scratch[1]);
+        C_ADD( scratch[3] , scratch[0] , scratch[2] );
+        C_SUB( scratch[4] , scratch[0] , scratch[2] );
+        C_SUB( Fout[m2], *Fout, scratch[3] );
+        tw1 += fstride;
+        tw2 += fstride*2;
+        tw3 += fstride*3;
+        C_ADDTO( *Fout , scratch[3] );
+
+        if(st->inverse) {
+            Fout[m].r = scratch[5].r - scratch[4].i;
+            Fout[m].i = scratch[5].i + scratch[4].r;
+            Fout[m3].r = scratch[5].r + scratch[4].i;
+            Fout[m3].i = scratch[5].i - scratch[4].r;
+        }else{
+            Fout[m].r = scratch[5].r + scratch[4].i;
+            Fout[m].i = scratch[5].i - scratch[4].r;
+            Fout[m3].r = scratch[5].r - scratch[4].i;
+            Fout[m3].i = scratch[5].i + scratch[4].r;
+        }
+        ++Fout;
+    }while(--k);
+}
+
+static void kf_bfly3(
+         kiss_fft_cpx * Fout,
+         const size_t fstride,
+         const kiss_fft_cfg st,
+         size_t m
+         )
+{
+     size_t k=m;
+     const size_t m2 = 2*m;
+     kiss_fft_cpx *tw1,*tw2;
+     kiss_fft_cpx scratch[5];
+     kiss_fft_cpx epi3;
+     epi3 = st->twiddles[fstride*m];
+
+     tw1=tw2=st->twiddles;
+
+     do{
+         C_FIXDIV(*Fout,3); C_FIXDIV(Fout[m],3); C_FIXDIV(Fout[m2],3);
+
+         C_MUL(scratch[1],Fout[m] , *tw1);
+         C_MUL(scratch[2],Fout[m2] , *tw2);
+
+         C_ADD(scratch[3],scratch[1],scratch[2]);
+         C_SUB(scratch[0],scratch[1],scratch[2]);
+         tw1 += fstride;
+         tw2 += fstride*2;
+
+         Fout[m].r = Fout->r - HALF_OF(scratch[3].r);
+         Fout[m].i = Fout->i - HALF_OF(scratch[3].i);
+
+         C_MULBYSCALAR( scratch[0] , epi3.i );
+
+         C_ADDTO(*Fout,scratch[3]);
+
+         Fout[m2].r = Fout[m].r + scratch[0].i;
+         Fout[m2].i = Fout[m].i - scratch[0].r;
+
+         Fout[m].r -= scratch[0].i;
+         Fout[m].i += scratch[0].r;
+
+         ++Fout;
+     }while(--k);
+}
+
+static void kf_bfly5(
+        kiss_fft_cpx * Fout,
+        const size_t fstride,
+        const kiss_fft_cfg st,
+        int m
+        )
+{
+    kiss_fft_cpx *Fout0,*Fout1,*Fout2,*Fout3,*Fout4;
+    int u;
+    kiss_fft_cpx scratch[13];
+    kiss_fft_cpx * twiddles = st->twiddles;
+    kiss_fft_cpx *tw;
+    kiss_fft_cpx ya,yb;
+    ya = twiddles[fstride*m];
+    yb = twiddles[fstride*2*m];
+
+    Fout0=Fout;
+    Fout1=Fout0+m;
+    Fout2=Fout0+2*m;
+    Fout3=Fout0+3*m;
+    Fout4=Fout0+4*m;
+
+    tw=st->twiddles;
+    for ( u=0; u<m; ++u ) {
+        C_FIXDIV( *Fout0,5); C_FIXDIV( *Fout1,5); C_FIXDIV( *Fout2,5); C_FIXDIV( *Fout3,5); C_FIXDIV( *Fout4,5);
+        scratch[0] = *Fout0;
+
+        C_MUL(scratch[1] ,*Fout1, tw[u*fstride]);
+        C_MUL(scratch[2] ,*Fout2, tw[2*u*fstride]);
+        C_MUL(scratch[3] ,*Fout3, tw[3*u*fstride]);
+        C_MUL(scratch[4] ,*Fout4, tw[4*u*fstride]);
+
+        C_ADD( scratch[7],scratch[1],scratch[4]);
+        C_SUB( scratch[10],scratch[1],scratch[4]);
+        C_ADD( scratch[8],scratch[2],scratch[3]);
+        C_SUB( scratch[9],scratch[2],scratch[3]);
+
+        Fout0->r += scratch[7].r + scratch[8].r;
+        Fout0->i += scratch[7].i + scratch[8].i;
+
+        scratch[5].r = scratch[0].r + S_MUL(scratch[7].r,ya.r) + S_MUL(scratch[8].r,yb.r);
+        scratch[5].i = scratch[0].i + S_MUL(scratch[7].i,ya.r) + S_MUL(scratch[8].i,yb.r);
+
+        scratch[6].r =  S_MUL(scratch[10].i,ya.i) + S_MUL(scratch[9].i,yb.i);
+        scratch[6].i = -S_MUL(scratch[10].r,ya.i) - S_MUL(scratch[9].r,yb.i);
+
+        C_SUB(*Fout1,scratch[5],scratch[6]);
+        C_ADD(*Fout4,scratch[5],scratch[6]);
+
+        scratch[11].r = scratch[0].r + S_MUL(scratch[7].r,yb.r) + S_MUL(scratch[8].r,ya.r);
+        scratch[11].i = scratch[0].i + S_MUL(scratch[7].i,yb.r) + S_MUL(scratch[8].i,ya.r);
+        scratch[12].r = - S_MUL(scratch[10].i,yb.i) + S_MUL(scratch[9].i,ya.i);
+        scratch[12].i = S_MUL(scratch[10].r,yb.i) - S_MUL(scratch[9].r,ya.i);
+
+        C_ADD(*Fout2,scratch[11],scratch[12]);
+        C_SUB(*Fout3,scratch[11],scratch[12]);
+
+        ++Fout0;++Fout1;++Fout2;++Fout3;++Fout4;
+    }
+}
+
+/* perform the butterfly for one stage of a mixed radix FFT */
+static void kf_bfly_generic(
+        kiss_fft_cpx * Fout,
+        const size_t fstride,
+        const kiss_fft_cfg st,
+        int m,
+        int p
+        )
+{
+    int u,k,q1,q;
+    kiss_fft_cpx * twiddles = st->twiddles;
+    kiss_fft_cpx t;
+    int Norig = st->nfft;
+
+    kiss_fft_cpx * scratch = (kiss_fft_cpx*)KISS_FFT_TMP_ALLOC(sizeof(kiss_fft_cpx)*p);
+
+    for ( u=0; u<m; ++u ) {
+        k=u;
+        for ( q1=0 ; q1<p ; ++q1 ) {
+            scratch[q1] = Fout[ k  ];
+            C_FIXDIV(scratch[q1],p);
+            k += m;
+        }
+
+        k=u;
+        for ( q1=0 ; q1<p ; ++q1 ) {
+            int twidx=0;
+            Fout[ k ] = scratch[0];
+            for (q=1;q<p;++q ) {
+                twidx += fstride * k;
+                if (twidx>=Norig) twidx-=Norig;
+                C_MUL(t,scratch[q] , twiddles[twidx] );
+                C_ADDTO( Fout[ k ] ,t);
+            }
+            k += m;
+        }
+    }
+    KISS_FFT_TMP_FREE(scratch);
+}
+
+static
+void kf_work(
+        kiss_fft_cpx * Fout,
+        const kiss_fft_cpx * f,
+        const size_t fstride,
+        int in_stride,
+        int * factors,
+        const kiss_fft_cfg st
+        )
+{
+    kiss_fft_cpx * Fout_beg=Fout;
+    const int p=*factors++; /* the radix  */
+    const int m=*factors++; /* stage's fft length/p */
+    const kiss_fft_cpx * Fout_end = Fout + p*m;
+
+#ifdef _OPENMP
+    // use openmp extensions at the 
+    // top-level (not recursive)
+    if (fstride==1 && p<=5)
+    {
+        int k;
+
+        // execute the p different work units in different threads
+#       pragma omp parallel for
+        for (k=0;k<p;++k) 
+            kf_work( Fout +k*m, f+ fstride*in_stride*k,fstride*p,in_stride,factors,st);
+        // all threads have joined by this point
+
+        switch (p) {
+            case 2: kf_bfly2(Fout,fstride,st,m); break;
+            case 3: kf_bfly3(Fout,fstride,st,m); break; 
+            case 4: kf_bfly4(Fout,fstride,st,m); break;
+            case 5: kf_bfly5(Fout,fstride,st,m); break; 
+            default: kf_bfly_generic(Fout,fstride,st,m,p); break;
+        }
+        return;
+    }
+#endif
+
+    if (m==1) {
+        do{
+            *Fout = *f;
+            f += fstride*in_stride;
+        }while(++Fout != Fout_end );
+    }else{
+        do{
+            // recursive call:
+            // DFT of size m*p performed by doing
+            // p instances of smaller DFTs of size m, 
+            // each one takes a decimated version of the input
+            kf_work( Fout , f, fstride*p, in_stride, factors,st);
+            f += fstride*in_stride;
+        }while( (Fout += m) != Fout_end );
+    }
+
+    Fout=Fout_beg;
+
+    // recombine the p smaller DFTs 
+    switch (p) {
+        case 2: kf_bfly2(Fout,fstride,st,m); break;
+        case 3: kf_bfly3(Fout,fstride,st,m); break; 
+        case 4: kf_bfly4(Fout,fstride,st,m); break;
+        case 5: kf_bfly5(Fout,fstride,st,m); break; 
+        default: kf_bfly_generic(Fout,fstride,st,m,p); break;
+    }
+}
+
+/*  facbuf is populated by p1,m1,p2,m2, ...
+    where 
+    p[i] * m[i] = m[i-1]
+    m0 = n                  */
+static 
+void kf_factor(int n,int * facbuf)
+{
+    int p=4;
+    double floor_sqrt;
+    floor_sqrt = floor( sqrt((double)n) );
+
+    /*factor out powers of 4, powers of 2, then any remaining primes */
+    do {
+        while (n % p) {
+            switch (p) {
+                case 4: p = 2; break;
+                case 2: p = 3; break;
+                default: p += 2; break;
+            }
+            if (p > floor_sqrt)
+                p = n;          /* no more factors, skip to end */
+        }
+        n /= p;
+        *facbuf++ = p;
+        *facbuf++ = n;
+    } while (n > 1);
+}
+
+/*
+ *
+ * User-callable function to allocate all necessary storage space for the fft.
+ *
+ * The return value is a contiguous block of memory, allocated with malloc.  As such,
+ * It can be freed with free(), rather than a kiss_fft-specific function.
+ * */
+kiss_fft_cfg kiss_fft_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem )
+{
+    kiss_fft_cfg st=NULL;
+    size_t memneeded = sizeof(struct kiss_fft_state)
+        + sizeof(kiss_fft_cpx)*(nfft-1); /* twiddle factors*/
+
+    if ( lenmem==NULL ) {
+        st = ( kiss_fft_cfg)KISS_FFT_MALLOC( memneeded );
+    }else{
+        if (mem != NULL && *lenmem >= memneeded)
+            st = (kiss_fft_cfg)mem;
+        *lenmem = memneeded;
+    }
+    if (st) {
+        int i;
+        st->nfft=nfft;
+        st->inverse = inverse_fft;
+
+        for (i=0;i<nfft;++i) {
+            const double pi=3.141592653589793238462643383279502884197169399375105820974944;
+            double phase = -2*pi*i / nfft;
+            if (st->inverse)
+                phase *= -1;
+            kf_cexp(st->twiddles+i, phase );
+        }
+
+        kf_factor(nfft,st->factors);
+    }
+    return st;
+}
+
+
+void kiss_fft_stride(kiss_fft_cfg st,const kiss_fft_cpx *fin,kiss_fft_cpx *fout,int in_stride)
+{
+    if (fin == fout) {
+        //NOTE: this is not really an in-place FFT algorithm.
+        //It just performs an out-of-place FFT into a temp buffer
+        kiss_fft_cpx * tmpbuf = (kiss_fft_cpx*)KISS_FFT_TMP_ALLOC( sizeof(kiss_fft_cpx)*st->nfft);
+        kf_work(tmpbuf,fin,1,in_stride, st->factors,st);
+        memcpy(fout,tmpbuf,sizeof(kiss_fft_cpx)*st->nfft);
+        KISS_FFT_TMP_FREE(tmpbuf);
+    }else{
+        kf_work( fout, fin, 1,in_stride, st->factors,st );
+    }
+}
+
+void kiss_fft(kiss_fft_cfg cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout)
+{
+    kiss_fft_stride(cfg,fin,fout,1);
+}
+
+
+void kiss_fft_cleanup(void)
+{
+    // nothing needed any more
+}
+
+int kiss_fft_next_fast_size(int n)
+{
+    while(1) {
+        int m=n;
+        while ( (m%2) == 0 ) m/=2;
+        while ( (m%3) == 0 ) m/=3;
+        while ( (m%5) == 0 ) m/=5;
+        if (m<=1)
+            break; /* n is completely factorable by twos, threes, and fives */
+        n++;
+    }
+    return n;
+}

--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/src/vamp-sdk/ext/kiss_fft.h	Tue Aug 16 16:04:09 2016 +0100
@@ -0,0 +1,124 @@
+#ifndef KISS_FFT_H
+#define KISS_FFT_H
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <math.h>
+#include <string.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*
+ ATTENTION!
+ If you would like a :
+ -- a utility that will handle the caching of fft objects
+ -- real-only (no imaginary time component ) FFT
+ -- a multi-dimensional FFT
+ -- a command-line utility to perform ffts
+ -- a command-line utility to perform fast-convolution filtering
+
+ Then see kfc.h kiss_fftr.h kiss_fftnd.h fftutil.c kiss_fastfir.c
+  in the tools/ directory.
+*/
+
+#ifdef USE_SIMD
+# include <xmmintrin.h>
+# define kiss_fft_scalar __m128
+#define KISS_FFT_MALLOC(nbytes) _mm_malloc(nbytes,16)
+#define KISS_FFT_FREE _mm_free
+#else	
+#define KISS_FFT_MALLOC malloc
+#define KISS_FFT_FREE free
+#endif	
+
+
+#ifdef FIXED_POINT
+#include <sys/types.h>	
+# if (FIXED_POINT == 32)
+#  define kiss_fft_scalar int32_t
+# else	
+#  define kiss_fft_scalar int16_t
+# endif
+#else
+# ifndef kiss_fft_scalar
+/*  default is float */
+#   define kiss_fft_scalar float
+# endif
+#endif
+
+typedef struct {
+    kiss_fft_scalar r;
+    kiss_fft_scalar i;
+}kiss_fft_cpx;
+
+typedef struct kiss_fft_state* kiss_fft_cfg;
+
+/* 
+ *  kiss_fft_alloc
+ *  
+ *  Initialize a FFT (or IFFT) algorithm's cfg/state buffer.
+ *
+ *  typical usage:      kiss_fft_cfg mycfg=kiss_fft_alloc(1024,0,NULL,NULL);
+ *
+ *  The return value from fft_alloc is a cfg buffer used internally
+ *  by the fft routine or NULL.
+ *
+ *  If lenmem is NULL, then kiss_fft_alloc will allocate a cfg buffer using malloc.
+ *  The returned value should be free()d when done to avoid memory leaks.
+ *  
+ *  The state can be placed in a user supplied buffer 'mem':
+ *  If lenmem is not NULL and mem is not NULL and *lenmem is large enough,
+ *      then the function places the cfg in mem and the size used in *lenmem
+ *      and returns mem.
+ *  
+ *  If lenmem is not NULL and ( mem is NULL or *lenmem is not large enough),
+ *      then the function returns NULL and places the minimum cfg 
+ *      buffer size in *lenmem.
+ * */
+
+kiss_fft_cfg kiss_fft_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem); 
+
+/*
+ * kiss_fft(cfg,in_out_buf)
+ *
+ * Perform an FFT on a complex input buffer.
+ * for a forward FFT,
+ * fin should be  f[0] , f[1] , ... ,f[nfft-1]
+ * fout will be   F[0] , F[1] , ... ,F[nfft-1]
+ * Note that each element is complex and can be accessed like
+    f[k].r and f[k].i
+ * */
+void kiss_fft(kiss_fft_cfg cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout);
+
+/*
+ A more generic version of the above function. It reads its input from every Nth sample.
+ * */
+void kiss_fft_stride(kiss_fft_cfg cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout,int fin_stride);
+
+/* If kiss_fft_alloc allocated a buffer, it is one contiguous 
+   buffer and can be simply free()d when no longer needed*/
+#define kiss_fft_free free
+
+/*
+ Cleans up some memory that gets managed internally. Not necessary to call, but it might clean up 
+ your compiler output to call this before you exit.
+*/
+void kiss_fft_cleanup(void);
+	
+
+/*
+ * Returns the smallest integer k, such that k>=n and k has only "fast" factors (2,3,5)
+ */
+int kiss_fft_next_fast_size(int n);
+
+/* for real ffts, we need an even size */
+#define kiss_fftr_next_fast_size_real(n) \
+        (kiss_fft_next_fast_size( ((n)+1)>>1)<<1)
+
+#ifdef __cplusplus
+} 
+#endif
+
+#endif

--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/src/vamp-sdk/ext/kiss_fftr.c	Tue Aug 16 16:04:09 2016 +0100
@@ -0,0 +1,159 @@
+/*
+Copyright (c) 2003-2004, Mark Borgerding
+
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
+    * Neither the author nor the names of any contributors may be used to endorse or promote products derived from this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#include "kiss_fftr.h"
+#include "_kiss_fft_guts.h"
+
+struct kiss_fftr_state{
+    kiss_fft_cfg substate;
+    kiss_fft_cpx * tmpbuf;
+    kiss_fft_cpx * super_twiddles;
+#ifdef USE_SIMD
+    void * pad;
+#endif
+};
+
+kiss_fftr_cfg kiss_fftr_alloc(int nfft,int inverse_fft,void * mem,size_t * lenmem)
+{
+    int i;
+    kiss_fftr_cfg st = NULL;
+    size_t subsize, memneeded;
+
+    if (nfft & 1) {
+        fprintf(stderr,"Real FFT optimization must be even.\n");
+        return NULL;
+    }
+    nfft >>= 1;
+
+    kiss_fft_alloc (nfft, inverse_fft, NULL, &subsize);
+    memneeded = sizeof(struct kiss_fftr_state) + subsize + sizeof(kiss_fft_cpx) * ( nfft * 3 / 2);
+
+    if (lenmem == NULL) {
+        st = (kiss_fftr_cfg) KISS_FFT_MALLOC (memneeded);
+    } else {
+        if (*lenmem >= memneeded)
+            st = (kiss_fftr_cfg) mem;
+        *lenmem = memneeded;
+    }
+    if (!st)
+        return NULL;
+
+    st->substate = (kiss_fft_cfg) (st + 1); /*just beyond kiss_fftr_state struct */
+    st->tmpbuf = (kiss_fft_cpx *) (((char *) st->substate) + subsize);
+    st->super_twiddles = st->tmpbuf + nfft;
+    kiss_fft_alloc(nfft, inverse_fft, st->substate, &subsize);
+
+    for (i = 0; i < nfft/2; ++i) {
+        double phase =
+            -3.14159265358979323846264338327 * ((double) (i+1) / nfft + .5);
+        if (inverse_fft)
+            phase *= -1;
+        kf_cexp (st->super_twiddles+i,phase);
+    }
+    return st;
+}
+
+void kiss_fftr(kiss_fftr_cfg st,const kiss_fft_scalar *timedata,kiss_fft_cpx *freqdata)
+{
+    /* input buffer timedata is stored row-wise */
+    int k,ncfft;
+    kiss_fft_cpx fpnk,fpk,f1k,f2k,tw,tdc;
+
+    if ( st->substate->inverse) {
+        fprintf(stderr,"kiss fft usage error: improper alloc\n");
+        exit(1);
+    }
+
+    ncfft = st->substate->nfft;
+
+    /*perform the parallel fft of two real signals packed in real,imag*/
+    kiss_fft( st->substate , (const kiss_fft_cpx*)timedata, st->tmpbuf );
+    /* The real part of the DC element of the frequency spectrum in st->tmpbuf
+     * contains the sum of the even-numbered elements of the input time sequence
+     * The imag part is the sum of the odd-numbered elements
+     *
+     * The sum of tdc.r and tdc.i is the sum of the input time sequence. 
+     *      yielding DC of input time sequence
+     * The difference of tdc.r - tdc.i is the sum of the input (dot product) [1,-1,1,-1... 
+     *      yielding Nyquist bin of input time sequence
+     */
+ 
+    tdc.r = st->tmpbuf[0].r;
+    tdc.i = st->tmpbuf[0].i;
+    C_FIXDIV(tdc,2);
+    CHECK_OVERFLOW_OP(tdc.r ,+, tdc.i);
+    CHECK_OVERFLOW_OP(tdc.r ,-, tdc.i);
+    freqdata[0].r = tdc.r + tdc.i;
+    freqdata[ncfft].r = tdc.r - tdc.i;
+#ifdef USE_SIMD    
+    freqdata[ncfft].i = freqdata[0].i = _mm_set1_ps(0);
+#else
+    freqdata[ncfft].i = freqdata[0].i = 0;
+#endif
+
+    for ( k=1;k <= ncfft/2 ; ++k ) {
+        fpk    = st->tmpbuf[k]; 
+        fpnk.r =   st->tmpbuf[ncfft-k].r;
+        fpnk.i = - st->tmpbuf[ncfft-k].i;
+        C_FIXDIV(fpk,2);
+        C_FIXDIV(fpnk,2);
+
+        C_ADD( f1k, fpk , fpnk );
+        C_SUB( f2k, fpk , fpnk );
+        C_MUL( tw , f2k , st->super_twiddles[k-1]);
+
+        freqdata[k].r = HALF_OF(f1k.r + tw.r);
+        freqdata[k].i = HALF_OF(f1k.i + tw.i);
+        freqdata[ncfft-k].r = HALF_OF(f1k.r - tw.r);
+        freqdata[ncfft-k].i = HALF_OF(tw.i - f1k.i);
+    }
+}
+
+void kiss_fftri(kiss_fftr_cfg st,const kiss_fft_cpx *freqdata,kiss_fft_scalar *timedata)
+{
+    /* input buffer timedata is stored row-wise */
+    int k, ncfft;
+
+    if (st->substate->inverse == 0) {
+        fprintf (stderr, "kiss fft usage error: improper alloc\n");
+        exit (1);
+    }
+
+    ncfft = st->substate->nfft;
+
+    st->tmpbuf[0].r = freqdata[0].r + freqdata[ncfft].r;
+    st->tmpbuf[0].i = freqdata[0].r - freqdata[ncfft].r;
+    C_FIXDIV(st->tmpbuf[0],2);
+
+    for (k = 1; k <= ncfft / 2; ++k) {
+        kiss_fft_cpx fk, fnkc, fek, fok, tmp;
+        fk = freqdata[k];
+        fnkc.r = freqdata[ncfft - k].r;
+        fnkc.i = -freqdata[ncfft - k].i;
+        C_FIXDIV( fk , 2 );
+        C_FIXDIV( fnkc , 2 );
+
+        C_ADD (fek, fk, fnkc);
+        C_SUB (tmp, fk, fnkc);
+        C_MUL (fok, tmp, st->super_twiddles[k-1]);
+        C_ADD (st->tmpbuf[k],     fek, fok);
+        C_SUB (st->tmpbuf[ncfft - k], fek, fok);
+#ifdef USE_SIMD        
+        st->tmpbuf[ncfft - k].i *= _mm_set1_ps(-1.0);
+#else
+        st->tmpbuf[ncfft - k].i *= -1;
+#endif
+    }
+    kiss_fft (st->substate, st->tmpbuf, (kiss_fft_cpx *) timedata);
+}

--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/src/vamp-sdk/ext/kiss_fftr.h	Tue Aug 16 16:04:09 2016 +0100
@@ -0,0 +1,46 @@
+#ifndef KISS_FTR_H
+#define KISS_FTR_H
+
+#include "kiss_fft.h"
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+    
+/* 
+ 
+ Real optimized version can save about 45% cpu time vs. complex fft of a real seq.
+
+ 
+ 
+ */
+
+typedef struct kiss_fftr_state *kiss_fftr_cfg;
+
+
+kiss_fftr_cfg kiss_fftr_alloc(int nfft,int inverse_fft,void * mem, size_t * lenmem);
+/*
+ nfft must be even
+
+ If you don't care to allocate space, use mem = lenmem = NULL 
+*/
+
+
+void kiss_fftr(kiss_fftr_cfg cfg,const kiss_fft_scalar *timedata,kiss_fft_cpx *freqdata);
+/*
+ input timedata has nfft scalar points
+ output freqdata has nfft/2+1 complex points
+*/
+
+void kiss_fftri(kiss_fftr_cfg cfg,const kiss_fft_cpx *freqdata,kiss_fft_scalar *timedata);
+/*
+ input freqdata has  nfft/2+1 complex points
+ output timedata has nfft scalar points
+*/
+
+#define kiss_fftr_free free
+
+#ifdef __cplusplus
+}
+#endif
+#endif

--- a/vamp-sdk/FFT.h	Wed May 18 12:21:17 2016 +0100
+++ b/vamp-sdk/FFT.h	Tue Aug 16 16:04:09 2016 +0100
@@ -44,22 +44,19 @@
 
 /**
  * A simple FFT implementation provided for convenience of plugin
- * authors.
- * 
- * This class provides double-precision FFTs in power-of-two sizes
- * only. It is slower than more sophisticated library
- * implementations. If these requirements aren't suitable, make other
- * arrangements.
+ * authors. This class provides one-shot (i.e. fixed table state is
+ * recalculated every time) double precision complex-complex
+ * transforms. For repeated transforms from real time-domain data, use
+ * a FFTReal object instead.
  *
- * The inverse transform is scaled by 1/n.
- *
- * The implementation is from Don Cross's public domain FFT code.
+ * The forward transform is unscaled; the inverse transform is scaled
+ * by 1/n.
  */
 class FFT
 {
 public:
     /**
-     * Calculate a forward transform of size n.
+     * Calculate a one-shot forward transform of size n.
      * n must be a power of 2, greater than 1.
      *
      * ri and ii must point to the real and imaginary component arrays
@@ -75,7 +72,7 @@
                         double *ro, double *io);
 
     /**
-     * Calculate an inverse transform of size n.
+     * Calculate a one-shot inverse transform of size n.
      * n must be a power of 2, greater than 1.
      *
      * ri and ii must point to the real and imaginary component arrays
@@ -93,6 +90,51 @@
                         double *ro, double *io);
 };
 
+/**
+ * A simple FFT implementation provided for convenience of plugin
+ * authors. This class provides transforms between real
+ * single-precision time-domain and complex single-precision
+ * frequency-domain data.
+ *
+ * The forward transform is unscaled; the inverse transform is scaled
+ * by 1/n.
+ */
+class FFTReal
+{
+    /**
+     * Prepare to calculate transforms of size n.
+     * n must be a power of 2, greater than 1.
+     */
+    FFTReal(unsigned int n);
+
+    ~FFTReal();
+
+    /**
+     * Calculate a forward transform of size n.
+     *
+     * ri must point to the real input data of size n.
+     *
+     * co must point to enough space to receive an interleaved complex
+     * output array of size n/2+1.
+     */
+    void forward(const float *ri, float *co);
+
+    /**
+     * Calculate an inverse transform of size n.
+     *
+     * ci must point to an interleaved complex input array of size n/2+1.
+     *
+     * ro must point to enough space to receive the real output data
+     * of size n. The output is scaled by 1/n and only the real part
+     * is returned.
+     */
+    void inverse(const float *ci, float *ro);
+
+private:
+    class D;
+    D *m_d;
+};
+
 }
 
 _VAMP_SDK_PLUGSPACE_END(FFT.h)