Generative Fast Fourier Transforms (GFFT)  0.3
Generative Fast Fourier Transforms (GFFT) Documentation

The main feature of GFFT library is that the length of transform is defined as a compile-time constant. You may compile FFT of only single length or easily define and compile a set of transforms, e.g. the powers of 2 from 1 to 25 or a range 10,...,20, etc.

The major advantage is that the first and expensive planning step usually performed by the known FFT libraries is fully done at compile-time. Even the twiddle factors are partially computed during compilation. Knowing the length of transform, the compiler can orginize and optimize the code better. As result, at run-time GFFT performs the transform in one step avoiding expensive planning.

GFFT is a header-only library with a couple of sample programs and some files from Loki-lib library by Andrei Alexandrescu. It doesn't need any installation or additional packages.

What's new in the version 0.3

  1. Out-of-place transforms. In the version 0.2, the only in-place transforms were possible.
  2. Complex transforms of any N. Although, transforms for prime factors are computed as DFT without recursion. Therefore, N with big prime factors will be significantly slower.
  3. Better accuracy due to compile-time computation of the roots-of-unity as rationals. Their precision is 18 decimal digits.
  4. The syntax changed not significantly: GenerateTransform template class receives the list of transform lengths as the first parameter instead of the min. and max. powers of 2. The last new parameter should define either IN_PLACE or OUT_OF_PLACE
  5. Real transforms remain unchanged, only in-place of the size 2^N. They will be reimplemented in the next release according to the new architecture of the complex transforms.

Compilation

Starting with version 0.2 GFFT has cmake-based build management. That means, you can take a look and make changes in CMakeList.txt file, where all the compiler options and targets are listed.
To compile the project, run cmake in GFFT directory, then run make:

cmake .
make

Compilation of GFFT is compiler-challenging process, because it intensively applies template class recursion, which must be completely resolved during compilation. For instance, it is known that gcc 3.x hangs up, when trying to compile GFFT in optimized mode. Newer compilers can handle template classes much better and faster. GFFT 0.3 is succesfully tested with

Other compilers will be tested soon.

The compilation should take from a few seconds to a few minutes depending on amount of transforms you would like to compile within your code.

Basic usage

To start using GFFT please take a look into simple example programs (gfft*.cpp and cgfft*.cpp). The first one defines the data as a C-like array, where each even element represents the real and odd - the imaginary part of a complex number. The second one uses array of std::complex type.

Since all the parameters of GFFT are static constants and defined as template parameters, you have to deside before compilation, which kind of transforms of which length you might need and declare them in instantiation of template class GFFT::GenerateTransform (see its documentation for details). Its template parameters may be given as a single type or as a typelist with multiple options.

An object of this instantiated template class contains then object factory of all the needed transforms. Each of them can be obtained from the object factory on demand. Following example declares forward and backward complex transform of double precision in single-threaded mode and creates a transform object for forward transform of length 2^10:

using namespace GFFT;
typedef TYPELIST_2(DFT, IDFT) ComplexTransforms;
typedef TYPELIST_3(SIntID<8>,SIntID<16>,SIntID<32>) NList;
typedef GenerateTransform<NList, DOUBLE, ComplexTransforms, SIntID<1>, Serial, OUT_OF_PLACE> TransformSet;
TransformSet gfft;
TransformSet::ObjectType* fftobj = gfft.CreateTransformObject(10, DOUBLE::ID, DFT::ID, 1, Serial::ID, OUT_OF_PLACE::ID);

If you need only single transform type of fixed length, then you can use directly template class GFFT::Transform without object factory.


Generated on Mon Feb 10 2014 for Generative Fast Fourier Transforms (GFFT) by DoxyGen 1.8.3.1