first addition of kissfft.hh the C++ template fft engine

This commit is contained in:
Mark Borgerding 2009-05-17 23:57:26 +00:00
parent 9dbaf860f2
commit 2b5477d54c
3 changed files with 407 additions and 0 deletions

325
kissfft.hh Normal file
View File

@ -0,0 +1,325 @@
#ifndef KISSFFT_CLASS_HH
#include <complex>
#include <vector>
namespace kissfft_utils {
template <class T_twid>
struct traits
{
void fill_twiddles( std::complex<T_twid> * dst ,int nfft,bool inverse);
void prepare(
std::vector< std::complex<T_twid> > & dst,
int nfft,bool inverse,
std::vector<int> & stageRadix,
std::vector<int> & stageRemainder )
{
dst.resize(nfft);
fill_twiddles( &dst[0],nfft,inverse);
//factorize
//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;// no more factors
}
n /= p;
stageRadix.push_back(p);
stageRemainder.push_back(n);
}while(n>1);
}
};
template <class T_twid>
void traits<T_twid>::fill_twiddles( std::complex<T_twid> * dst ,int nfft,bool inverse)
{
T_twid phinc = (inverse?2:-2)* acos( (T_twid) -1) / nfft;
for (int i=0;i<nfft;++i)
dst[i] = exp( std::complex<T_twid>(0,i*phinc) );
}
/*
template <>
void traits<long double>::fill_twiddles(std::complex<long double> * dst ,int nfft,bool inverse)
{
long double phinc = (inverse?2:-2)*3.14159265358979323846264338327950288419716939937510L / nfft;
for (int i=0;i<nfft;++i)
dst[i] = std::complex<long double>(cosl(i*phinc),sinl(i*phinc));
}
*/
}
template <typename T_Data,typename T_traits=kissfft_utils::traits<T_Data> >
class kissfft
{
public:
typedef T_traits traits_type;
typedef T_Data scalar_type;
typedef std::complex<scalar_type> cpx_type;
kissfft(int nfft,bool inverse,const traits_type & traits=traits_type() )
:_nfft(nfft),_inverse(inverse),_traits(traits)
{
_traits.prepare(_twiddles, _nfft,_inverse ,_stageRadix, _stageRemainder);
}
void transform(const cpx_type * src , cpx_type * dst)
{
kf_work(0, dst, src, 1);
}
private:
void kf_work( int stage,cpx_type * Fout, const cpx_type * f, const size_t fstride)
{
int p = _stageRadix[stage];
int m = _stageRemainder[stage];
cpx_type * Fout_beg = Fout;
cpx_type * Fout_end = Fout + p*m;
if (m==1) {
do{
*Fout = *f;
f += fstride;
}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(stage+1, Fout , f, fstride*p);
f += fstride;
}while( (Fout += m) != Fout_end );
}
Fout=Fout_beg;
// recombine the p smaller DFTs
switch (p) {
case 2: kf_bfly2(Fout,fstride,m); break;
case 3: kf_bfly3(Fout,fstride,m); break;
case 4: kf_bfly4(Fout,fstride,m); break;
case 5: kf_bfly5(Fout,fstride,m); break;
default: kf_bfly_generic(Fout,fstride,m,p); break;
}
}
// these were #define macros in the original kiss_fft
void C_ADD( cpx_type & c,const cpx_type & a,const cpx_type & b) { c=a+b;}
void C_MUL( cpx_type & c,const cpx_type & a,const cpx_type & b) { c=a*b;}
void C_SUB( cpx_type & c,const cpx_type & a,const cpx_type & b) { c=a-b;}
void C_ADDTO( cpx_type & c,const cpx_type & a) { c+=a;}
void C_FIXDIV( cpx_type & c,int n) {} // NO-OP for float types
scalar_type S_MUL( const scalar_type & a,const scalar_type & b) { return a*b;}
scalar_type HALF_OF( const scalar_type & a) { return a*.5;}
void C_MULBYSCALAR(cpx_type & c,const scalar_type & a) {c*=a;}
void kf_bfly2( cpx_type * Fout, const size_t fstride, int m)
{
cpx_type * Fout2;
cpx_type * tw1 = &_twiddles[0];
cpx_type 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);
}
void kf_bfly4( cpx_type * Fout, const size_t fstride, const size_t m)
{
cpx_type *tw1,*tw2,*tw3;
cpx_type scratch[6];
size_t k=m;
const size_t m2=2*m;
const size_t m3=3*m;
tw3 = tw2 = tw1 = &_twiddles[0];
do {
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(_inverse) {
Fout[m] = cpx_type( scratch[5].real() - scratch[4].imag() , scratch[5].imag() + scratch[4].real() );
Fout[m3] = cpx_type( scratch[5].real() + scratch[4].imag() , scratch[5].imag() - scratch[4].real() );
}else{
Fout[m] = cpx_type( scratch[5].real() + scratch[4].imag() , scratch[5].imag() - scratch[4].real() );
Fout[m3] = cpx_type( scratch[5].real() - scratch[4].imag() , scratch[5].imag() + scratch[4].real() );
}
++Fout;
}while(--k);
}
void kf_bfly3( cpx_type * Fout, const size_t fstride, const size_t m)
{
size_t k=m;
const size_t m2 = 2*m;
cpx_type *tw1,*tw2;
cpx_type scratch[5];
cpx_type epi3;
epi3 = _twiddles[fstride*m];
tw1=tw2=&_twiddles[0];
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] = cpx_type( Fout->real() - HALF_OF(scratch[3].real() ) , Fout->imag() - HALF_OF(scratch[3].imag() ) );
C_MULBYSCALAR( scratch[0] , epi3.imag() );
C_ADDTO(*Fout,scratch[3]);
Fout[m2] = cpx_type( Fout[m].real() + scratch[0].imag() , Fout[m].imag() - scratch[0].real() );
C_ADDTO( Fout[m] , cpx_type( -scratch[0].imag(),scratch[0].real() ) );
++Fout;
}while(--k);
}
void kf_bfly5( cpx_type * Fout, const size_t fstride, const size_t m)
{
cpx_type *Fout0,*Fout1,*Fout2,*Fout3,*Fout4;
int u;
cpx_type scratch[13];
cpx_type * twiddles = &_twiddles[0];
cpx_type *tw;
cpx_type 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 ) {
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]);
C_ADDTO( *Fout0, scratch[7]);
C_ADDTO( *Fout0, scratch[8]);
scratch[5] = scratch[0] + cpx_type(
S_MUL(scratch[7].real(),ya.real() ) + S_MUL(scratch[8].real() ,yb.real() ),
S_MUL(scratch[7].imag(),ya.real()) + S_MUL(scratch[8].imag(),yb.real())
);
scratch[6] = cpx_type(
S_MUL(scratch[10].imag(),ya.imag()) + S_MUL(scratch[9].imag(),yb.imag()),
-S_MUL(scratch[10].real(),ya.imag()) - S_MUL(scratch[9].real(),yb.imag())
);
C_SUB(*Fout1,scratch[5],scratch[6]);
C_ADD(*Fout4,scratch[5],scratch[6]);
scratch[11] = scratch[0] +
cpx_type(
S_MUL(scratch[7].real(),yb.real()) + S_MUL(scratch[8].real(),ya.real()),
S_MUL(scratch[7].imag(),yb.real()) + S_MUL(scratch[8].imag(),ya.real())
);
scratch[12] = cpx_type(
-S_MUL(scratch[10].imag(),yb.imag()) + S_MUL(scratch[9].imag(),ya.imag()),
S_MUL(scratch[10].real(),yb.imag()) - S_MUL(scratch[9].real(),ya.imag())
);
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 */
void kf_bfly_generic(
cpx_type * Fout,
const size_t fstride,
int m,
int p
)
{
int u,k,q1,q;
cpx_type * twiddles = &_twiddles[0];
cpx_type t;
int Norig = _nfft;
cpx_type scratchbuf[p];
for ( u=0; u<m; ++u ) {
k=u;
for ( q1=0 ; q1<p ; ++q1 ) {
scratchbuf[q1] = Fout[ k ];
C_FIXDIV(scratchbuf[q1],p);
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;
C_MUL(t,scratchbuf[q] , twiddles[twidx] );
C_ADDTO( Fout[ k ] ,t);
}
k += m;
}
}
}
int _nfft;
bool _inverse;
std::vector<cpx_type> _twiddles;
std::vector<int> _stageRadix;
std::vector<int> _stageRemainder;
traits_type _traits;
};
#endif

View File

@ -95,5 +95,11 @@ selftest.c:
selftest_short.c:
./mk_test.py -s 10 12 14 > selftest_short.c
CXXFLAGS=-O3 -ffast-math -fomit-frame-pointer -I.. -I../tools -W -Wall
testcpp: testcpp.cc ../kissfft.hh
$(CXX) -o $@ $(CXXFLAGS) -lm testcpp.cc
clean:
rm -f *~ bm_* st_* tr_* kf_* tkfc_* ff_* ffr_* *.pyc *.pyo *.dat

76
test/testcpp.cc Normal file
View File

@ -0,0 +1,76 @@
#include "kissfft.hh"
#include <iostream>
#include <cstdlib>
#include <typeinfo>
#include <sys/time.h>
static inline
double curtime(void)
{
struct timeval tv;
gettimeofday(&tv, NULL);
return (double)tv.tv_sec + (double)tv.tv_usec*.000001;
}
using namespace std;
template <class T>
void dotest(int nfft)
{
typedef kissfft<T> FFT;
typedef std::complex<T> cpx_type;
cout << typeid(T).name() << ":nfft=" << nfft <<endl;
FFT fft(nfft,false);
vector<cpx_type> inbuf(nfft);
vector<cpx_type> outbuf(nfft);
#if 0
for (int k=0;k<nfft;++k)
inbuf[k]= cpx_type(
cosl(2*k* M_PIl / nfft ),
sinl(2*k* M_PIl / nfft ) );
#else
for (int k=0;k<nfft;++k)
inbuf[k]= cpx_type(
(T)(rand()/(double)RAND_MAX - .5),
(T)(rand()/(double)RAND_MAX - .5) );
#endif
fft.transform( &inbuf[0] , &outbuf[0] );
long double totalpower=0;
long double difpower=0;
for (int k0=0;k0<nfft;++k0) {
complex<long double> acc = 0;
long double phinc = 2*k0* M_PIl / nfft;
for (int k1=0;k1<nfft;++k1) {
complex<long double> x(inbuf[k1].real(),inbuf[k1].imag());
acc += x * exp( complex<long double>(0,-k1*phinc) );
}
totalpower += norm(acc);
complex<long double> x(outbuf[k0].real(),outbuf[k0].imag());
complex<long double> dif = acc - x;
difpower += norm(dif);
}
cout << "RMSE:" << sqrt(difpower/totalpower) << "\t";
double t0 = curtime();
int nits=20e6/nfft;
for (int k=0;k<nits;++k) {
fft.transform( &inbuf[0] , &outbuf[0] );
}
double t1 = curtime();
cout << "MSPS:" << ( (nits*nfft)*1e-6/ (t1-t0) ) << endl;
}
int main(int argc,char ** argv)
{
dotest<float>(32);
dotest<double>(32);
dotest<long double>(32);
dotest<float>(1024); dotest<double>(1024); dotest<long double>(1024);
dotest<float>(1800); dotest<double>(1800); dotest<long double>(1800);
return 0;
}