FFTServer.h

Go to the documentation of this file.

//# FFTServer.h: A class with methods for Fast Fourier Transforms
//# Copyright (C) 1994,1995,1996,1997,1999,2003
//# Associated Universities, Inc. Washington DC, USA.
//#
//# This library is free software; you can redistribute it and/or modify it
//# under the terms of the GNU Library General Public License as published by
//# the Free Software Foundation; either version 2 of the License, or (at your
//# option) any later version.
//#
//# This library is distributed in the hope that it will be useful, but WITHOUT
//# ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
//# FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Library General Public
//# License for more details.
//#
//# You should have received a copy of the GNU Library General Public License
//# along with this library; if not, write to the Free Software Foundation,
//# Inc., 675 Massachusetts Ave, Cambridge, MA 02139, USA.
//#
//# Correspondence concerning AIPS++ should be addressed as follows:
//#        Internet email: aips2-request@nrao.edu.
//#        Postal address: AIPS++ Project Office
//#                        National Radio Astronomy Observatory
//#                        520 Edgemont Road
//#                        Charlottesville, VA 22903-2475 USA
//#
//# $Id$

#ifndef SCIMATH_FFTSERVER_H
#define SCIMATH_FFTSERVER_H

//# Includes
#include <casacore/casa/aips.h>
#include <casacore/scimath/Mathematics/FFTW.h>
#include <casacore/casa/Arrays/IPosition.h>
#include <casacore/casa/Containers/Block.h>
#include <vector>

namespace casacore { //# NAMESPACE CASACORE - BEGIN

//# Forward Declarations
template <class T> class Array;
template <class S> class Matrix;
// <summary>Lists the different types of FFT's that can be done</summary>
// <synopsis>This enumerator is brought out as a separate class because g++
// currently cannot handle enumerators in a templated class. When it can this
// class will go away and this enumerator moved into the FFTServer
// class</synopsis>
class FFTEnums {
public:
  enum TransformType {
    // Forward Complex to Complex transforms.
    COMPLEX,
    // Inverse Complex to Complex transforms.
    INVCOMPLEX,
    // Real to Complex or Complex to Real transforms.
    REALTOCOMPLEX,
    // Real to Complex or Complex to Real transforms.
    COMPLEXTOREAL,
    // Real to Real transforms with symmetric Arrays (not used)
    REALSYMMETRIC
  };
};

// <summary>A class with methods for Fast Fourier Transforms</summary>

// <use visibility=export>

// <reviewed reviewer="wbrouw" date="1997/10/29" tests="tFFTServer">
// </reviewed>

// <prerequisite> 
// <li> Basic concepts of Fast Fourier Transforms.
// <li> <linkto module=Arrays>The Arrays module</linkto>
// </prerequisite>

// <etymology> The FFTServer class, can do Fast Fourier Transforms of
// any length and dimensionality.
// </etymology>


// <synopsis>

// The FFTServer class provides methods for performing n-dimensional Fast
// Fourier Transforms with real and complex Array's of arbitrary size and
// dimensionality. It can do either real to complex, complex to real, or
// complex to complex transforms with the "origin" of the transform either at
// the centre of the Array or at the first element.

// Because the output from a real to complex transform is Hermitian only half
// of the complex result is returned. Similarly with a complex to real
// transform only half of the complex plane is required, the other half is
// implicitly assumed to be the complex conjugate of the supplied half-plane.
// <note role=warning> The complex to real transform does not check that the
// imaginary component of the values where u=0 are zero</note>

// This class can be initialised with a shape that indicates the length of the
// transforms that will be performed, and whether they are going to be
// real<->complex transforms or complex<->complex ones. This initialisation
// sets up a variety of internal buffers and computes factorizations and
// twiddle factors used during the transform. The initialised transform shape
// is always compared with the shape of the supplied arguments when a transform
// is done and the FFTServer class will automatically resize itself if
// necessary. So the default constructor is perfectly safe to use.

// With any transform the output Array must either be the correct shape for the
// desired output or zero length (ie not contain any elements). If it is zero
// length then it will be resized to the correct shape. For a complex->complex
// transform the output Array will be the same shape as the input Array. For a
// real->complex transform the output Array will be the same size as the input
// Array except in the first dimension which will have a length of (nx+2)/2. So
// if nx=7 the output length will be 4 and if nx=8 the output length will be 5,
// on the first axis. nx is the length of the first axis on the <em>real</em>
// input Array and cx (which is used later) is the length of the first axis on
// the <em>complex</em> input Array.

// <strong>For complex to real transforms the output length on the first axis
// is not uniquely defined by the shape of the complex input
// Array</strong>. This class uses the following algorithm to work out the
// length of the first axis on the output Array.
// <ul> 
// <li> If the size of the output Array is non-zero then its shape must match
// the size of the input Array except for the first axis. The length of the
// first axis must either be 2*cx-2 or 2*cx-1 and this determines the length of
// the transform on the first axis.
// <li> If the size of the output Array is zero then scan the imaginary
// components of the values at the end of the first axis on the input Array (ie
// at <src>[cx-1,....]</src> If any of these are non-zero the output Array
// will have an odd length.
// <li> Otherwise if all the imaginary components described above are zero then
// look at the current size of the FFTServer object (either defined at
// construction time or with the resize function). If it matches the size of
// the input Array except for the first axis and if the length on this axis is
// either 2*cx-2 or 2*cx-1 then use that to determine the size of the output
// Array.
// <li> Otherwise assume the output Array will an even length of 2*cx-2 on its
// first axis.
// </ul>

// This class does transforms using the widely used FORTRAN fftpack
// package or the highly optimized FFTW package (to be chosen at build time).
// <br>
// <em> P.N. Swarztrauber, Vectorizing the FFTs, in Parallel Computations
// (G. Rodrigue, ed.), Academic Press, 1982, pp. 51--83. </em><br>
// The fftpack package only does one dimensional transforms and this class
// decomposes multi-dimensional transforms into a series of 1-dimensional ones.
// <br>If at build time it is chosen to use FFTW in a multi-threaded way,
// it will try to use as many cores as possible.

// In this class a forward transform is defined as one that goes from the real
// to the complex (or the time to frequency) domain. In a forward transform the
// sign of the exponent is negative and no scaling is done on the output.  The
// backward transform goes from the complex to the real (or the frequency to
// the time) domain. The sign of the exponent is positive and the result is
// always scaled by 1/N were N is the total number of elements in the Array.

// The origin of the transform is defined as the point where setting only that
// element to one, and then doing a forward transform results in an Array that
// is all one. The <src>fft</src> member functions in this class all assume
// that the origin of the Transform is at the centre of the Array ie. at
// <src>[nx/2,ny/2,...]</src> were the indexing begins at zero. Because the
// fftpack software assumes the origin of the transform is at the first element
// ie.,<src>[0,0,...]</src> this class flips the data in the Array around to
// compensate. For fftpack this flipping takes about one 20% of the total
// transform time, while for FFTW it can easily exceed the transform time.
// Flipping can be avoided by using the <src>fft0</src> member
// functions which do not flip the data.

// Some of the member functions in this class scramble the input Array,
// possibly by flipping the quandrants of the data although this is not
// guaranteed. Modification of the input Array can be avoided, at the expense
// of copying the data to temporary storage, by either:
// <ul> <li> Ensuring the input Array is a const Array.
//      <li> Setting the constInput Flag to True.
// </ul>
// The latter option is provided to avoid users having to cast non-const
// Arrays to const ones in order to prevent there input Array from being
// scrambled.

// <note role=warning> This class assumes that a Complex array is stored as
// pairs of floating point numbers, with no intervening gaps, and with the real
// component first ie., <src>[re0,im0,re1,im1, ...]</src>. This means that the
// following type casts work,
// <srcblock>
// S * complexPtr;
// T * realPtr = (T * ) complexPtr;
// </srcblock>
// and allow a Complex number to be accessed as a pair of real numbers. If this
// assumption is bad then real Arrays will have to generated by copying the
// complex ones. Ultimately this assumption about Complex<->Real Array
// conversion should be put somewhere central like Array2Math.cc.
// </note>
// </synopsis>

// <templating arg=T>
// <li> The T argument must be of type Float or Double. These are the only 
// possible instantiations of this class.
// </templating>

// <templating arg=S>
// <li> The S argument must be of type Complex, if T is Float, or DComplex, if T is
// Double. These are the only possible instantiations of this class.
// </templating>
//
// <example>
// Do a real to complex Transform of a 1-Dimensional Vector.  The following
// example can trivially be extended to any number of dimensions.
// <srcblock>
// FFTServer<Float,Complex> server;
// Vector<Float> input(32);
// Vector<Complex> output(17);
// input = 0.0f;
// input(16) = 1.0f;
// cout << "Input:" << input << endl;
// server.fft(output, input);
// cout << "Output:" << output << endl;
// </srcblock>
// </example>
//
// <thrown>
// <li> AipsError: If the input and output Array have bad or incompatible
// shapes. See the individual function descriptions for what Array shapes are
// required.
// </thrown>
//
// <todo asof="1997/10/22">
//   <li> The time taken to flip the Array can be reduced, if all the Array
//   dimensions are even, by pre-multiplying the every other element on the
//   input Array by -1. Then no flipping needs to be done on the output Array.
// </todo>

template<class T, class S> class FFTServer
{
public:

  // The default constructor. The server will automatically resize to do
  // transforms of the appropriate length when necessary.
  FFTServer();

  // Initialise the server to do transforms on Arrays of the specified
  // shape. The server will, however, resize to do transforms of other lengths
  // if necessary. See the resize function for a description of the
  // TransformType enumerator.
  FFTServer(const IPosition & fftSize, 
            const FFTEnums::TransformType transformType 
            = FFTEnums::REALTOCOMPLEX);
  
  // copy constructor. The copied server is initialised to do transforms of the
  // same length as the other server. Uses copy (and not reference) semantics
  // so that changing the transform length of one server does not affect the
  // other server.
  FFTServer(const FFTServer<T,S> & other);

  // destructor
  ~FFTServer();
  
  // The assignment operator which does the same thing as the copy
  // constructor.
  FFTServer<T,S> & operator=(const FFTServer<T,S> & other);

  // Modify the FFTServer object to do transforms of the supplied shape. The
  // amount of internal storage, and the initialisation, depends on the type of
  // transform that will be done. The transform type is specified with the
  // TransformTypes enumerator. Currently there is no difference in
  // initialisation for the COMPLEXTOREAL and REALTOCOMPLEX transforms. The
  // shape argument is the shape of the real array (or complex one if complex
  // to complex transforms are being done). In general it is not necessary to
  // use this function as all the fft & fft0 functions will automatically
  // resize the server, if necessary, to match their input arguments.
  void resize(const IPosition & fftSize,
              const FFTEnums::TransformType transformType
              = FFTEnums::REALTOCOMPLEX);

  // Real to complex fft. The origin of the transform is in the centre of the
  // Array. Because of the Hermitian property the output Array only contains
  // half of the complex result. The output Array must either have no elements
  // or be a size that is appropriate to the input Array size,
  // ie. <src>shape = [(nx+2)/2, ny, nz,...]</src>.  Otherwise an AipsError is
  // thrown. See the synopsis for a description of the constInput flag.
  // <group>
  void fft(Array<S> & cResult, Array<T> & rData, const Bool constInput=False);
  void fft(Array<S> & cResult, const Array<T> & rData);
  // </group>

  // Complex to real fft. The origin of the transform is in the centre of the
  // Array. Because of the Hermitian property the input Array only contains
  // half of the complex values. The output Array must either have no elements,
  // or be a size that is appropriate to the input Array size ie.,<br>
  // <src>shape = [2*cx-2, cy, cz,...]</src> or <br>
  // <src>shape = [2*cx-1, cy, cz,...]</src>.  <br>
  // Otherwise an AipsError is thrown. See the description in the synopsis for
  // the algorithm used to choose between the two possible output shapes and a
  // description of the constInput Flag.  
  // <group>
  void fft(Array<T> & rResult, Array<S> & cData, const Bool constInput=False);
  void fft(Array<T> & rResult, const Array<S> & cData);
  // </group>

  // Complex to complex in-place fft. The origin of the transform is in the
  // centre of the Array. The direction of the transform is controlled by the
  // toFrequency variable. If True then a forward, or time to frequency,
  // transform is performed. If False a backward or frequency to time transform
  // is done. Scaling is always done on the backward transform.
  void fft(Array<S> & cValues, const Bool toFrequency=True);

  // Complex to complex fft. The origin of the transform is in the centre of
  // the Array. The direction of the transform is controlled by the toFrequency
  // variable. If True then a forward, or time to frequency, transform is
  // performed. If False a backward or frequency to time transform is
  // done. Scaling is always done on the backward transform. The output Array
  // must either either contain no elements or be the same as the input Array,
  // ie. <src>shape = [cx, cy, cz,...]</src>.  Otherwise an AipsError is
  // thrown.
  void fft(Array<S> & cResult, const Array<S> & cData,
           const Bool toFrequency=True);

  // The <src>fft0</src> functions are equivalent to the <src>fft</src>
  // functions described above except that the origin of the transform is the
  // first element of the Array, ie. [0,0,0...], rather than the centre
  // element, ie [nx/2, ny/2, nz/2, ...]. As the underlying functions
  // assume that the origin of the transform is the first element these
  // routines are in general faster than the equivalent ones with the origin
  // at the centre of the Array.
  // <group>
  void fft0(Array<S> & cResult, Array<T> & rData, const Bool constInput=False);
  void fft0(Array<S> & cResult, const Array<T> & rData);
  void fft0(Array<T> & rResult, Array<S> & cData, const Bool constInput=False);
  void fft0(Array<T> & rResult, const Array<S> & cData);
  void fft0(Array<S> & cValues, const Bool toFrequency=True);
  void fft0(Array<S> & cResult, const Array<S> & cData,
            const Bool toFrequency=True);
  //# void fft0(Array<T> & rValues, const Bool toFrequency=True);

  // </group>
  //# Flips the quadrants in a complex Array so that the point at
  //# cData.shape()/2 moves to the origin. This moves, for example, the point
  //# at [8,3] to the origin ([0,0]) in an array of shape [16,7]. Usually two
  //# flips will restore an Array to its original state.  But for Array's
  //# where one or more dimension is an odd length two flips do NOT restore
  //# the data to its original state.  So the when toZero=False this routine
  //# does an unflip operation (ie moves the data at [0,0] to the centre) and
  //# restores the data to its original state for odd length arrays.  When
  //# passed a Hermitian Array where half the complex plane is implicit (eg as
  //# produced by a real->complex Transform) it is not necessary to flip the
  //# first dimension of the Array. In this case the isHermitian flag should
  //# be set to True.  For complex<->complex transforms this should be False.
  // <group>
  void flip(Array<T> & rData, const Bool toZero, const Bool isHermitian);
  void flip(Array<S> & cData, const Bool toZero, const Bool isHermitian);
  // </group>

  // N-D in-place complex->complex FFT shift (FFT - phase-mult - inverse FFT)
  // If toFrequency is true, the first FFT will be from time to frequency. 
  // relshift is the freq shift normalised to the bandwidth.
  // Only transform over selected dimension. Iterate over the others. 
  void fftshift(Array<S> & cValues, const uInt& whichAxis, 
                const Double& relshift, const Bool toFrequency=True);

  // N-D complex->complex FFT shift (FFT - phase-mult - inverse FFT)
  // with flagging.
  // If toFrequency is true, the first FFT will be from time to frequency. 
  // relshift is the freq shift normalised to the bandwidth.
  // Only transform over selected dimension. Iterate over the others. 
  void fftshift(Array<S> & outValues, Array<Bool> & outFlags,
                const Array<S> & cValues, const Array<Bool>& inFlags,
                const uInt& whichAxis, 
                const Double& relshift, 
                const Bool goodIsTrue=False,
                const Bool toFrequency=True);

  // N-D real->real FFT shift (FFT to complex - phase-mult - inverse FFT)
  // with flagging.
  // relshift is the freq shift normalised to the bandwidth.
  // Only transform over selected dimension. Iterate over the others. 
  void fftshift(Array<T> & outValues, Array<Bool> & outFlags,
                const Array<T> & rValues, const Array<Bool>& inFlags,
                const uInt& whichAxis, 
                const Double& relshift, 
                const Bool goodIsTrue=False);

private:
  //# finds the shape of the output array when doing complex->real transforms
  IPosition determineShape(const IPosition & rShape, const Array<S> & cData);

  //# Data members.
  // The size of the last FFT done by this object
  IPosition itsSize;
  // Whether the last FFT was complex<->complex or not
  FFTEnums::TransformType itsTransformType;
  //# twiddle factors and factorisations used by fftpack
  PtrBlock<Block<T> *> itsWork;
  // buffer for copying non-contigious arrays to contigious ones. This is done
  // so that the FFT's have a better chance of fitting into cache and hence
  // going faster. 
  // This buffer is also used as temporary storage when flipping the data.
  Block<S> itsBuffer;
  // FFTW specific members. Do not harm if FFTPack is used.
  FFTW           itsFFTW;
  std::vector<T> itsWorkIn;
  std::vector<S> itsWorkOut;
  std::vector<S> itsWorkC2C;
};


} //# NAMESPACE CASACORE - END

//# Do NOT include the .tcc file here like done for other templated classes.
//# The instantiations are done explicitly.
//# In this way the HAVE_FFTW ifdef is only used in .cc files and does
//# not appear in headers, so other packages using FFTServer do not need
//# to (un)set HAVE_FFTW.

#endif