diff options
author | Mark Borgerding <mark@borgerding.net> | 2009-05-27 21:32:42 -0400 |
---|---|---|
committer | Mark Borgerding <mark@borgerding.net> | 2009-05-27 21:32:42 -0400 |
commit | f13e000b454fed0a9b0d71ca1481d6894bacd8eb (patch) | |
tree | 5b12909ce991142f60143a4a1195859d5feec8e0 /unsupported/Eigen | |
parent | 09b47332553a79dab30516e6b1d410dea90cf9b7 (diff) |
various comment changes
Diffstat (limited to 'unsupported/Eigen')
-rw-r--r-- | unsupported/Eigen/FFT | 6 | ||||
-rw-r--r-- | unsupported/Eigen/src/FFT/ei_kissfft_impl.h | 484 |
2 files changed, 237 insertions, 253 deletions
diff --git a/unsupported/Eigen/FFT b/unsupported/Eigen/FFT index 3d852f5a2..31d8c74c5 100644 --- a/unsupported/Eigen/FFT +++ b/unsupported/Eigen/FFT @@ -1,5 +1,5 @@ // 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 // @@ -29,14 +29,14 @@ #include "src/FFT/ei_kissfft_impl.h" #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 // TODO // #include "src/FFT/ei_fftw_impl.h" // #define DEFAULT_FFT_IMPL ei_fftw_impl #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 // TODO // #include "src/FFT/ei_imkl_impl.h" diff --git a/unsupported/Eigen/src/FFT/ei_kissfft_impl.h b/unsupported/Eigen/src/FFT/ei_kissfft_impl.h index 453c7f6da..91fa5ca18 100644 --- a/unsupported/Eigen/src/FFT/ei_kissfft_impl.h +++ b/unsupported/Eigen/src/FFT/ei_kissfft_impl.h @@ -1,5 +1,5 @@ // 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 // @@ -28,252 +28,255 @@ namespace Eigen { + // This FFT implementation was derived from kissfft http:sourceforge.net/projects/kissfft + // Copyright 2003-2009 Mark Borgerding + template <typename _Scalar> - 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; - m_twiddles.resize(nfft); - 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 conjugate() + struct ei_kiss_cpx_fft { - m_inverse = !m_inverse; - for ( size_t i=0;i<m_twiddles.size() ;++i) - m_twiddles[i] = conj( m_twiddles[i] ); - } + 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 factorize(int nfft) - { - //start factoring out 4's, then 2's, then 3,5,7,9,... - int n= nfft; - int p=4; - do { - while (n % p) { + void make_twiddles(int nfft,bool inverse) + { + m_inverse = inverse; + m_twiddles.resize(nfft); + 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 conjugate() + { + m_inverse = !m_inverse; + for ( size_t i=0;i<m_twiddles.size() ;++i) + m_twiddles[i] = conj( m_twiddles[i] ); + } + + void factorize(int nfft) + { + //start factoring out 4's, then 2's, then 3,5,7,9,... + 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 4: p = 2; break; - case 2: p = 3; break; - default: p += 2; break; + 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 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; + void bfly2( Complex * Fout, const size_t fstride, int m) + { + for (int k=0;k<m;++k) { + Complex t = Fout[m+k] * m_twiddles[k*fstride]; + Fout[m+k] = Fout[k] - t; + Fout[k] += t; + } } - } - void bfly2( Complex * Fout, const size_t fstride, int m) - { - for (int k=0;k<m;++k) { - Complex t = Fout[m+k] * m_twiddles[k*fstride]; - Fout[m+k] = Fout[k] - t; - Fout[k] += t; + 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 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]; + + do{ + 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 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]; - - do{ - 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()) + 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()) ); - scratch[12] = Complex( - -(scratch[10].imag()*yb.imag()) + (scratch[9].imag()*ya.imag()), - (scratch[10].real()*yb.imag()) - (scratch[9].real()*ya.imag()) - ); + *Fout1 = scratch[5] - scratch[6]; + *Fout4 = scratch[5] + scratch[6]; - *Fout2=scratch[11]+scratch[12]; - *Fout3=scratch[11]-scratch[12]; + 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()) + ); - ++Fout0;++Fout1;++Fout2;++Fout3;++Fout4; - } - } - - /* perform the butterfly for one stage of a mixed radix FFT */ - void bfly_generic( - Complex * Fout, - 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 ) { - k=u; - for ( q1=0 ; q1<p ; ++q1 ) { - scratchbuf[q1] = Fout[ k ]; - k += m; + scratch[12] = Complex( + -(scratch[10].imag()*yb.imag()) + (scratch[9].imag()*ya.imag()), + (scratch[10].real()*yb.imag()) - (scratch[9].real()*ya.imag()) + ); + + *Fout2=scratch[11]+scratch[12]; + *Fout3=scratch[11]-scratch[12]; + + ++Fout0;++Fout1;++Fout2;++Fout3;++Fout4; } + } - 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; + /* perform the butterfly for one stage of a mixed radix FFT */ + void bfly_generic( + Complex * Fout, + 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 ) { + 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> - struct ei_kissfft_impl - { + struct ei_kissfft_impl + { typedef _Scalar Scalar; typedef std::complex<Scalar> Complex; @@ -284,10 +287,10 @@ namespace Eigen { } template <typename _Src> - void fwd( Complex * dst,const _Src *src,int nfft) - { - get_plan(nfft,false).work(0, dst, src, 1,1); - } + void fwd( Complex * dst,const _Src *src,int nfft) + { + get_plan(nfft,false).work(0, dst, src, 1,1); + } // real-to-complex forward FFT // 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 std::map<int,PlanData> PlanMap; + PlanMap m_plans; std::map<int, std::vector<Complex> > m_realTwiddles; std::vector<Complex> m_scratchBuf; @@ -376,25 +378,7 @@ namespace Eigen { PlanData & get_plan(int nfft,bool inverse) { - /* TODO: figure out why this does not work (g++ 4.3.2) - * 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; - */ + // TODO look for PlanKey(nfft, ! inverse) and conjugate the twiddles PlanData & pd = m_plans[ PlanKey(nfft,inverse) ]; if ( pd.m_twiddles.size() == 0 ) { pd.make_twiddles(nfft,inverse); @@ -421,5 +405,5 @@ namespace Eigen { for (int k=0;k<n;++k) dst[k] *= s; } - }; + }; } |