dsp::OptimalFFT Class Reference

Chooses the optimal FFT length for Filterbank and/or Convolution. More...

#include <OptimalFFT.h>

Inheritance diagram for dsp::OptimalFFT:

Public Member Functions
void	set_simultaneous (bool flag)
	Set true when convolution is performed during filterbank synthesis.
unsigned	get_nfft (unsigned nfilt) const
virtual void	set_nchan (unsigned nchan)
	Set the number of channels into which the data will be divided.
virtual double	compute_cost (unsigned nfft, unsigned nfilt) const
	Return the time required to execute in microseconds.
virtual std::string	get_library (unsigned nfft)
	Get the name of the FFT library to use for the given FFT length.
Public Member Functions inherited from Reference::Able
	Able (const Able &)
Able &	operator= (const Able &)
unsigned	get_reference_count () const
Public Member Functions inherited from Reference::HeapTracked
	HeapTracked (const HeapTracked &)
HeapTracked &	operator= (const HeapTracked &)
bool	__is_on_heap () const

Static Public Attributes
static bool	verbose = false

Protected Member Functions
virtual FTransform::Bench *	new_bench () const
Protected Member Functions inherited from Reference::Able
Handle *	__reference (bool active) const
void	__dereference (bool auto_delete=true) const

Protected Attributes
Reference::To< FTransform::Bench >	bench
unsigned	nchan
bool	simultaneous

Additional Inherited Members
Static Public Member Functions inherited from Reference::Able
static size_t	get_instance_count ()
Static Public Member Functions inherited from Reference::HeapTracked
static void *	operator new (size_t size, void *ptr=0)
static void	operator delete (void *location, void *ptr)
static void	operator delete (void *location)
static size_t	get_heap_queue_size ()

Detailed Description

Chooses the optimal FFT length for Filterbank and/or Convolution.

Member Function Documentation

get_nfft()

unsigned dsp::OptimalFFT::get_nfft

(

unsigned

nfilt

)

const

This method computes the optimal FFT length to use when performing multiple FFTs on a long time series. If FFT cost measurements are not available, this function assumes that the FFT is an O(NlogN) operation.

GIVEN: nchan - number of channels into which data will be divided nfilt - filter length (points discarded from each FFT in each channel)

RETURNS: The return value is the optimal FFT length, nfft

---

Where ngood = nfft - nfilt, the following ideas are held:

1) performance is better if ngood is a large fraction of nfft, 2) FFT performance is much better with smaller nfft.

The timescale for one FFT is proportional to NlogN, or:

           t = nfft * log(nfft)

The number of FFTs performed on a segment of M time samples is:

          Nf = M / (nfft - nfilt)

The total time spent on FFTs is then:

           T = t * Nf = nfft * log(nfft) * M/(nfft-nfilt)

Where M may be considered constant, and nfilt is given, this function aims to minimize:

     T(nfft) = nfft * log(nfft) / (nfft-nfilt)

---

The documentation for this class was generated from the following files:

• OptimalFFT.h
• OptimalFFT.C