various comment changes

This commit is contained in:
Mark Borgerding
2009-05-27 21:32:42 -04:00
parent 09b4733255
commit f13e000b45
2 changed files with 231 additions and 247 deletions

View File

@@ -1,5 +1,5 @@
// This file is part of Eigen, a lightweight C++ template library // This file is part of Eigen, a lightweight C++ template library
// for linear algebra. Eigen itself is part of the KDE project. // for linear algebra.
// //
// Copyright (C) 2009 Mark Borgerding mark a borgerding net // Copyright (C) 2009 Mark Borgerding mark a borgerding net
// //
@@ -29,14 +29,14 @@
#include "src/FFT/ei_kissfft_impl.h" #include "src/FFT/ei_kissfft_impl.h"
#define DEFAULT_FFT_IMPL ei_kissfft_impl #define DEFAULT_FFT_IMPL ei_kissfft_impl
// FFTW: faster, GPL-not LGPL, bigger code size // FFTW: faster, GPL -- incompatible with Eigen in LGPL form, bigger code size
#ifdef FFTW_PATIENT // definition of FFTW_PATIENT indicates the caller has included fftw3.h, we can use FFTW routines #ifdef FFTW_PATIENT // definition of FFTW_PATIENT indicates the caller has included fftw3.h, we can use FFTW routines
// TODO // TODO
// #include "src/FFT/ei_fftw_impl.h" // #include "src/FFT/ei_fftw_impl.h"
// #define DEFAULT_FFT_IMPL ei_fftw_impl // #define DEFAULT_FFT_IMPL ei_fftw_impl
#endif #endif
// intel Math Kernel Library: fastest, commerical // intel Math Kernel Library: fastest, commerical -- incompatible with Eigen in GPL form
#ifdef _MKL_DFTI_H_ // mkl_dfti.h has been included, we can use MKL FFT routines #ifdef _MKL_DFTI_H_ // mkl_dfti.h has been included, we can use MKL FFT routines
// TODO // TODO
// #include "src/FFT/ei_imkl_impl.h" // #include "src/FFT/ei_imkl_impl.h"

View File

@@ -1,5 +1,5 @@
// This file is part of Eigen, a lightweight C++ template library // This file is part of Eigen, a lightweight C++ template library
// for linear algebra. Eigen itself is part of the KDE project. // for linear algebra.
// //
// Copyright (C) 2009 Mark Borgerding mark a borgerding net // Copyright (C) 2009 Mark Borgerding mark a borgerding net
// //
@@ -28,252 +28,255 @@
namespace Eigen { namespace Eigen {
// This FFT implementation was derived from kissfft http:sourceforge.net/projects/kissfft
// Copyright 2003-2009 Mark Borgerding
template <typename _Scalar> template <typename _Scalar>
struct ei_kiss_cpx_fft struct ei_kiss_cpx_fft
{
typedef _Scalar Scalar;
typedef std::complex<Scalar> Complex;
std::vector<Complex> m_twiddles;
std::vector<int> m_stageRadix;
std::vector<int> m_stageRemainder;
bool m_inverse;
void make_twiddles(int nfft,bool inverse)
{ {
m_inverse = inverse; typedef _Scalar Scalar;
m_twiddles.resize(nfft); typedef std::complex<Scalar> Complex;
Scalar phinc = (inverse?2:-2)* acos( (Scalar) -1) / nfft; std::vector<Complex> m_twiddles;
for (int i=0;i<nfft;++i) std::vector<int> m_stageRadix;
m_twiddles[i] = exp( Complex(0,i*phinc) ); std::vector<int> m_stageRemainder;
} bool m_inverse;
void conjugate() void make_twiddles(int nfft,bool inverse)
{ {
m_inverse = !m_inverse; m_inverse = inverse;
for ( size_t i=0;i<m_twiddles.size() ;++i) m_twiddles.resize(nfft);
m_twiddles[i] = conj( m_twiddles[i] ); Scalar phinc = (inverse?2:-2)* acos( (Scalar) -1) / nfft;
} for (int i=0;i<nfft;++i)
m_twiddles[i] = exp( Complex(0,i*phinc) );
}
void factorize(int nfft) void conjugate()
{ {
//start factoring out 4's, then 2's, then 3,5,7,9,... m_inverse = !m_inverse;
int n= nfft; for ( size_t i=0;i<m_twiddles.size() ;++i)
int p=4; m_twiddles[i] = conj( m_twiddles[i] );
do { }
while (n % p) {
switch (p) { void factorize(int nfft)
case 4: p = 2; break; {
case 2: p = 3; break; //start factoring out 4's, then 2's, then 3,5,7,9,...
default: p += 2; break; int n= nfft;
int p=4;
do {
while (n % p) {
switch (p) {
case 4: p = 2; break;
case 2: p = 3; break;
default: p += 2; break;
}
if (p*p>n)
p=n;// impossible to have a factor > sqrt(n)
}
n /= p;
m_stageRadix.push_back(p);
m_stageRemainder.push_back(n);
}while(n>1);
}
template <typename _Src>
void work( int stage,Complex * xout, const _Src * xin, size_t fstride,size_t in_stride)
{
int p = m_stageRadix[stage];
int m = m_stageRemainder[stage];
Complex * Fout_beg = xout;
Complex * Fout_end = xout + p*m;
if (m>1) {
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
work(stage+1, xout , xin, fstride*p,in_stride);
xin += fstride*in_stride;
}while( (xout += m) != Fout_end );
}else{
do{
*xout = *xin;
xin += fstride*in_stride;
}while(++xout != Fout_end );
}
xout=Fout_beg;
// recombine the p smaller DFTs
switch (p) {
case 2: bfly2(xout,fstride,m); break;
case 3: bfly3(xout,fstride,m); break;
case 4: bfly4(xout,fstride,m); break;
case 5: bfly5(xout,fstride,m); break;
default: bfly_generic(xout,fstride,m,p); break;
} }
if (p*p>n)
p=n;// impossible to have a factor > sqrt(n)
} }
n /= p;
m_stageRadix.push_back(p);
m_stageRemainder.push_back(n);
}while(n>1);
}
template <typename _Src> void bfly2( Complex * Fout, const size_t fstride, int m)
void work( int stage,Complex * xout, const _Src * xin, size_t fstride,size_t in_stride) {
{ for (int k=0;k<m;++k) {
int p = m_stageRadix[stage]; Complex t = Fout[m+k] * m_twiddles[k*fstride];
int m = m_stageRemainder[stage]; Fout[m+k] = Fout[k] - t;
Complex * Fout_beg = xout; Fout[k] += t;
Complex * Fout_end = xout + p*m; }
}
void bfly4( Complex * Fout, const size_t fstride, const size_t m)
{
Complex scratch[6];
int negative_if_inverse = m_inverse * -2 +1;
for (size_t k=0;k<m;++k) {
scratch[0] = Fout[k+m] * m_twiddles[k*fstride];
scratch[1] = Fout[k+2*m] * m_twiddles[k*fstride*2];
scratch[2] = Fout[k+3*m] * m_twiddles[k*fstride*3];
scratch[5] = Fout[k] - scratch[1];
Fout[k] += scratch[1];
scratch[3] = scratch[0] + scratch[2];
scratch[4] = scratch[0] - scratch[2];
scratch[4] = Complex( scratch[4].imag()*negative_if_inverse , -scratch[4].real()* negative_if_inverse );
Fout[k+2*m] = Fout[k] - scratch[3];
Fout[k] += scratch[3];
Fout[k+m] = scratch[5] + scratch[4];
Fout[k+3*m] = scratch[5] - scratch[4];
}
}
void bfly3( Complex * Fout, const size_t fstride, const size_t m)
{
size_t k=m;
const size_t m2 = 2*m;
Complex *tw1,*tw2;
Complex scratch[5];
Complex epi3;
epi3 = m_twiddles[fstride*m];
tw1=tw2=&m_twiddles[0];
if (m>1) {
do{ do{
// recursive call: scratch[1]=Fout[m] * *tw1;
// DFT of size m*p performed by doing scratch[2]=Fout[m2] * *tw2;
// p instances of smaller DFTs of size m,
// each one takes a decimated version of the input scratch[3]=scratch[1]+scratch[2];
work(stage+1, xout , xin, fstride*p,in_stride); scratch[0]=scratch[1]-scratch[2];
xin += fstride*in_stride; tw1 += fstride;
}while( (xout += m) != Fout_end ); tw2 += fstride*2;
}else{ Fout[m] = Complex( Fout->real() - .5*scratch[3].real() , Fout->imag() - .5*scratch[3].imag() );
do{ scratch[0] *= epi3.imag();
*xout = *xin; *Fout += scratch[3];
xin += fstride*in_stride; Fout[m2] = Complex( Fout[m].real() + scratch[0].imag() , Fout[m].imag() - scratch[0].real() );
}while(++xout != Fout_end ); Fout[m] += Complex( -scratch[0].imag(),scratch[0].real() );
++Fout;
}while(--k);
} }
xout=Fout_beg;
// recombine the p smaller DFTs void bfly5( Complex * Fout, const size_t fstride, const size_t m)
switch (p) { {
case 2: bfly2(xout,fstride,m); break; Complex *Fout0,*Fout1,*Fout2,*Fout3,*Fout4;
case 3: bfly3(xout,fstride,m); break; size_t u;
case 4: bfly4(xout,fstride,m); break; Complex scratch[13];
case 5: bfly5(xout,fstride,m); break; Complex * twiddles = &m_twiddles[0];
default: bfly_generic(xout,fstride,m,p); break; Complex *tw;
} Complex ya,yb;
} ya = twiddles[fstride*m];
yb = twiddles[fstride*2*m];
void bfly2( Complex * Fout, const size_t fstride, int m) Fout0=Fout;
{ Fout1=Fout0+m;
for (int k=0;k<m;++k) { Fout2=Fout0+2*m;
Complex t = Fout[m+k] * m_twiddles[k*fstride]; Fout3=Fout0+3*m;
Fout[m+k] = Fout[k] - t; Fout4=Fout0+4*m;
Fout[k] += t;
}
}
void bfly4( Complex * Fout, const size_t fstride, const size_t m) tw=twiddles;
{ for ( u=0; u<m; ++u ) {
Complex scratch[6]; scratch[0] = *Fout0;
int negative_if_inverse = m_inverse * -2 +1;
for (size_t k=0;k<m;++k) {
scratch[0] = Fout[k+m] * m_twiddles[k*fstride];
scratch[1] = Fout[k+2*m] * m_twiddles[k*fstride*2];
scratch[2] = Fout[k+3*m] * m_twiddles[k*fstride*3];
scratch[5] = Fout[k] - scratch[1];
Fout[k] += scratch[1]; scratch[1] = *Fout1 * tw[u*fstride];
scratch[3] = scratch[0] + scratch[2]; scratch[2] = *Fout2 * tw[2*u*fstride];
scratch[4] = scratch[0] - scratch[2]; scratch[3] = *Fout3 * tw[3*u*fstride];
scratch[4] = Complex( scratch[4].imag()*negative_if_inverse , -scratch[4].real()* negative_if_inverse ); scratch[4] = *Fout4 * tw[4*u*fstride];
Fout[k+2*m] = Fout[k] - scratch[3]; scratch[7] = scratch[1] + scratch[4];
Fout[k] += scratch[3]; scratch[10] = scratch[1] - scratch[4];
Fout[k+m] = scratch[5] + scratch[4]; scratch[8] = scratch[2] + scratch[3];
Fout[k+3*m] = scratch[5] - scratch[4]; scratch[9] = scratch[2] - scratch[3];
}
}
void bfly3( Complex * Fout, const size_t fstride, const size_t m) *Fout0 += scratch[7];
{ *Fout0 += scratch[8];
size_t k=m;
const size_t m2 = 2*m;
Complex *tw1,*tw2;
Complex scratch[5];
Complex epi3;
epi3 = m_twiddles[fstride*m];
tw1=tw2=&m_twiddles[0]; scratch[5] = scratch[0] + Complex(
(scratch[7].real()*ya.real() ) + (scratch[8].real() *yb.real() ),
do{ (scratch[7].imag()*ya.real()) + (scratch[8].imag()*yb.real())
scratch[1]=Fout[m] * *tw1;
scratch[2]=Fout[m2] * *tw2;
scratch[3]=scratch[1]+scratch[2];
scratch[0]=scratch[1]-scratch[2];
tw1 += fstride;
tw2 += fstride*2;
Fout[m] = Complex( Fout->real() - .5*scratch[3].real() , Fout->imag() - .5*scratch[3].imag() );
scratch[0] *= epi3.imag();
*Fout += scratch[3];
Fout[m2] = Complex( Fout[m].real() + scratch[0].imag() , Fout[m].imag() - scratch[0].real() );
Fout[m] += Complex( -scratch[0].imag(),scratch[0].real() );
++Fout;
}while(--k);
}
void bfly5( Complex * Fout, const size_t fstride, const size_t m)
{
Complex *Fout0,*Fout1,*Fout2,*Fout3,*Fout4;
size_t u;
Complex scratch[13];
Complex * twiddles = &m_twiddles[0];
Complex *tw;
Complex 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=twiddles;
for ( u=0; u<m; ++u ) {
scratch[0] = *Fout0;
scratch[1] = *Fout1 * tw[u*fstride];
scratch[2] = *Fout2 * tw[2*u*fstride];
scratch[3] = *Fout3 * tw[3*u*fstride];
scratch[4] = *Fout4 * tw[4*u*fstride];
scratch[7] = scratch[1] + scratch[4];
scratch[10] = scratch[1] - scratch[4];
scratch[8] = scratch[2] + scratch[3];
scratch[9] = scratch[2] - scratch[3];
*Fout0 += scratch[7];
*Fout0 += scratch[8];
scratch[5] = scratch[0] + Complex(
(scratch[7].real()*ya.real() ) + (scratch[8].real() *yb.real() ),
(scratch[7].imag()*ya.real()) + (scratch[8].imag()*yb.real())
);
scratch[6] = Complex(
(scratch[10].imag()*ya.imag()) + (scratch[9].imag()*yb.imag()),
-(scratch[10].real()*ya.imag()) - (scratch[9].real()*yb.imag())
);
*Fout1 = scratch[5] - scratch[6];
*Fout4 = scratch[5] + scratch[6];
scratch[11] = scratch[0] +
Complex(
(scratch[7].real()*yb.real()) + (scratch[8].real()*ya.real()),
(scratch[7].imag()*yb.real()) + (scratch[8].imag()*ya.real())
); );
scratch[12] = Complex( scratch[6] = Complex(
-(scratch[10].imag()*yb.imag()) + (scratch[9].imag()*ya.imag()), (scratch[10].imag()*ya.imag()) + (scratch[9].imag()*yb.imag()),
(scratch[10].real()*yb.imag()) - (scratch[9].real()*ya.imag()) -(scratch[10].real()*ya.imag()) - (scratch[9].real()*yb.imag())
); );
*Fout2=scratch[11]+scratch[12]; *Fout1 = scratch[5] - scratch[6];
*Fout3=scratch[11]-scratch[12]; *Fout4 = scratch[5] + scratch[6];
++Fout0;++Fout1;++Fout2;++Fout3;++Fout4; scratch[11] = scratch[0] +
} Complex(
} (scratch[7].real()*yb.real()) + (scratch[8].real()*ya.real()),
(scratch[7].imag()*yb.real()) + (scratch[8].imag()*ya.real())
);
/* perform the butterfly for one stage of a mixed radix FFT */ scratch[12] = Complex(
void bfly_generic( -(scratch[10].imag()*yb.imag()) + (scratch[9].imag()*ya.imag()),
Complex * Fout, (scratch[10].real()*yb.imag()) - (scratch[9].real()*ya.imag())
const size_t fstride, );
int m,
int p
)
{
int u,k,q1,q;
Complex * twiddles = &m_twiddles[0];
Complex t;
int Norig = m_twiddles.size();
Complex * scratchbuf = (Complex*)alloca(p*sizeof(Complex) );
for ( u=0; u<m; ++u ) { *Fout2=scratch[11]+scratch[12];
k=u; *Fout3=scratch[11]-scratch[12];
for ( q1=0 ; q1<p ; ++q1 ) {
scratchbuf[q1] = Fout[ k ]; ++Fout0;++Fout1;++Fout2;++Fout3;++Fout4;
k += m;
} }
}
k=u; /* perform the butterfly for one stage of a mixed radix FFT */
for ( q1=0 ; q1<p ; ++q1 ) { void bfly_generic(
int twidx=0; Complex * Fout,
Fout[ k ] = scratchbuf[0]; const size_t fstride,
for (q=1;q<p;++q ) { int m,
twidx += fstride * k; int p
if (twidx>=Norig) twidx-=Norig; )
t=scratchbuf[q] * twiddles[twidx]; {
Fout[ k ] += t; int u,k,q1,q;
Complex * twiddles = &m_twiddles[0];
Complex t;
int Norig = m_twiddles.size();
Complex * scratchbuf = (Complex*)alloca(p*sizeof(Complex) );
for ( u=0; u<m; ++u ) {
k=u;
for ( q1=0 ; q1<p ; ++q1 ) {
scratchbuf[q1] = Fout[ k ];
k += m;
}
k=u;
for ( q1=0 ; q1<p ; ++q1 ) {
int twidx=0;
Fout[ k ] = scratchbuf[0];
for (q=1;q<p;++q ) {
twidx += fstride * k;
if (twidx>=Norig) twidx-=Norig;
t=scratchbuf[q] * twiddles[twidx];
Fout[ k ] += t;
}
k += m;
} }
k += m;
} }
} }
} };
};
template <typename _Scalar> template <typename _Scalar>
struct ei_kissfft_impl struct ei_kissfft_impl
{ {
typedef _Scalar Scalar; typedef _Scalar Scalar;
typedef std::complex<Scalar> Complex; typedef std::complex<Scalar> Complex;
@@ -284,10 +287,10 @@ namespace Eigen {
} }
template <typename _Src> template <typename _Src>
void fwd( Complex * dst,const _Src *src,int nfft) void fwd( Complex * dst,const _Src *src,int nfft)
{ {
get_plan(nfft,false).work(0, dst, src, 1,1); get_plan(nfft,false).work(0, dst, src, 1,1);
} }
// real-to-complex forward FFT // real-to-complex forward FFT
// perform two FFTs of src even and src odd // perform two FFTs of src even and src odd
@@ -363,11 +366,10 @@ namespace Eigen {
} }
} }
private: private:
typedef ei_kiss_cpx_fft<Scalar> PlanData; typedef ei_kiss_cpx_fft<Scalar> PlanData;
typedef std::map<int,PlanData> PlanMap; typedef std::map<int,PlanData> PlanMap;
PlanMap m_plans; PlanMap m_plans;
std::map<int, std::vector<Complex> > m_realTwiddles; std::map<int, std::vector<Complex> > m_realTwiddles;
std::vector<Complex> m_scratchBuf; std::vector<Complex> m_scratchBuf;
@@ -376,25 +378,7 @@ namespace Eigen {
PlanData & get_plan(int nfft,bool inverse) PlanData & get_plan(int nfft,bool inverse)
{ {
/* TODO: figure out why this does not work (g++ 4.3.2) // TODO look for PlanKey(nfft, ! inverse) and conjugate the twiddles
* for some reason this does not work
*
PlanMap::iterator it;
it = m_plans.find( PlanKey(nfft,inverse) );
if (it == m_plans.end() ) {
// create new entry
it = m_plans.insert( make_pair( PlanKey(nfft,inverse) , PlanData() ) );
MapIt it2 = m_plans.find( PlanKey(nfft,!inverse) );
if (it2 != m_plans.end() ) {
it->second = it2.second;
it->second.conjugate();
}else{
it->second.make_twiddles(nfft,inverse);
it->second.factorize(nfft);
}
}
return it->second;
*/
PlanData & pd = m_plans[ PlanKey(nfft,inverse) ]; PlanData & pd = m_plans[ PlanKey(nfft,inverse) ];
if ( pd.m_twiddles.size() == 0 ) { if ( pd.m_twiddles.size() == 0 ) {
pd.make_twiddles(nfft,inverse); pd.make_twiddles(nfft,inverse);
@@ -421,5 +405,5 @@ namespace Eigen {
for (int k=0;k<n;++k) for (int k=0;k<n;++k)
dst[k] *= s; dst[k] *= s;
} }
}; };
} }