This repository has been archived on 2020-04-08. You can view files and clone it, but cannot push or open issues or pull requests.
Files
Indoor/lib/simple_fft/fft_impl.hpp
2018-01-24 11:44:08 +01:00

512 lines
15 KiB
C++

#ifndef __SIMPLE_FFT__FFT_IMPL_HPP__
#define __SIMPLE_FFT__FFT_IMPL_HPP__
#include "fft_settings.h"
#include "error_handling.hpp"
#include <cstddef>
#include <math.h>
#include <vector>
using std::size_t;
#ifndef M_PI
#define M_PI 3.1415926535897932
#endif
namespace simple_fft {
namespace impl {
enum FFT_direction
{
FFT_FORWARD = 0,
FFT_BACKWARD
};
// checking whether the size of array dimension is power of 2
// via "complement and compare" method
inline bool isPowerOfTwo(const size_t num)
{
if ((num == 0) || !(num & (~num + 1)))
return false;
return true;
}
inline bool checkNumElements(const size_t num_elements, const char *& error_description)
{
using namespace error_handling;
if (!isPowerOfTwo(num_elements)) {
GetErrorDescription(EC_ONE_OF_DIMS_ISNT_POWER_OF_TWO, error_description);
return false;
}
return true;
}
template <class TComplexArray1D>
inline void scaleValues(TComplexArray1D & data, const size_t num_elements)
{
real_type mult = 1.0 / num_elements;
int num_elements_signed = static_cast<int>(num_elements);
#ifndef __clang__
#ifdef __USE_OPENMP
#pragma omp parallel for
#endif
#endif
for(int i = 0; i < num_elements_signed; ++i) {
#ifdef __USE_SQUARE_BRACKETS_FOR_ELEMENT_ACCESS_OPERATOR
data[i] *= mult;
#else
data(i) *= mult;
#endif
}
}
// NOTE: explicit template specialization for the case of std::vector<complex_type>
// because it is used in 2D and 3D FFT for both array classes with square and round
// brackets of element access operator; I need to guarantee that sub-FFT 1D will
// use square brackets for element access operator anyway. It is pretty ugly
// to duplicate the code but I haven't found more elegant solution.
template <>
inline void scaleValues<std::vector<complex_type> >(std::vector<complex_type> & data,
const size_t num_elements)
{
real_type mult = 1.0 / num_elements;
int num_elements_signed = static_cast<int>(num_elements);
#ifndef __clang__
#ifdef __USE_OPENMP
#pragma omp parallel for
#endif
#endif
for(int i = 0; i < num_elements_signed; ++i) {
data[i] *= mult;
}
}
template <class TComplexArray1D>
inline void bufferExchangeHelper(TComplexArray1D & data, const size_t index_from,
const size_t index_to, complex_type & buf)
{
#ifdef __USE_SQUARE_BRACKETS_FOR_ELEMENT_ACCESS_OPERATOR
buf = data[index_from];
data[index_from] = data[index_to];
data[index_to]= buf;
#else
buf = data(index_from);
data(index_from) = data(index_to);
data(index_to)= buf;
#endif
}
// NOTE: explicit template specialization for the case of std::vector<complex_type>
// because it is used in 2D and 3D FFT for both array classes with square and round
// brackets of element access operator; I need to guarantee that sub-FFT 1D will
// use square brackets for element access operator anyway. It is pretty ugly
// to duplicate the code but I haven't found more elegant solution.
template <>
inline void bufferExchangeHelper<std::vector<complex_type> >(std::vector<complex_type> & data,
const size_t index_from,
const size_t index_to,
complex_type & buf)
{
buf = data[index_from];
data[index_from] = data[index_to];
data[index_to]= buf;
}
template <class TComplexArray1D>
void rearrangeData(TComplexArray1D & data, const size_t num_elements)
{
complex_type buf;
size_t target_index = 0;
size_t bit_mask;
for (size_t i = 0; i < num_elements; ++i)
{
if (target_index > i)
{
bufferExchangeHelper(data, target_index, i, buf);
}
// Initialize the bit mask
bit_mask = num_elements;
// While bit is 1
while (target_index & (bit_mask >>= 1)) // bit_mask = bit_mask >> 1
{
// Drop bit:
// & is bitwise AND,
// ~ is bitwise NOT
target_index &= ~bit_mask; // target_index = target_index & (~bit_mask)
}
// | is bitwise OR
target_index |= bit_mask; // target_index = target_index | bit_mask
}
}
template <class TComplexArray1D>
inline void fftTransformHelper(TComplexArray1D & data, const size_t match,
const size_t k, complex_type & product,
const complex_type factor)
{
#ifdef __USE_SQUARE_BRACKETS_FOR_ELEMENT_ACCESS_OPERATOR
product = data[match] * factor;
data[match] = data[k] - product;
data[k] += product;
#else
product = data(match) * factor;
data(match) = data(k) - product;
data(k) += product;
#endif
}
// NOTE: explicit template specialization for the case of std::vector<complex_type>
// because it is used in 2D and 3D FFT for both array classes with square and round
// brackets of element access operator; I need to guarantee that sub-FFT 1D will
// use square brackets for element access operator anyway. It is pretty ugly
// to duplicate the code but I haven't found more elegant solution.
template <>
inline void fftTransformHelper<std::vector<complex_type> >(std::vector<complex_type> & data,
const size_t match,
const size_t k,
complex_type & product,
const complex_type factor)
{
product = data[match] * factor;
data[match] = data[k] - product;
data[k] += product;
}
template <class TComplexArray1D>
bool makeTransform(TComplexArray1D & data, const size_t num_elements,
const FFT_direction fft_direction, const char *& error_description)
{
using namespace error_handling;
using std::sin;
double local_pi;
switch(fft_direction)
{
case(FFT_FORWARD):
local_pi = -M_PI;
break;
case(FFT_BACKWARD):
local_pi = M_PI;
break;
default:
GetErrorDescription(EC_WRONG_FFT_DIRECTION, error_description);
return false;
}
// declare variables to cycle the bits of initial signal
size_t next, match;
real_type sine;
real_type delta;
complex_type mult, factor, product;
// NOTE: user's complex type should have constructor like
// "complex(real, imag)", where each of real and imag has
// real type.
// cycle for all bit positions of initial signal
for (size_t i = 1; i < num_elements; i <<= 1)
{
next = i << 1; // getting the next bit
delta = local_pi / i; // angle increasing
sine = sin(0.5 * delta); // supplementary sin
// multiplier for trigonometric recurrence
mult = complex_type(-2.0 * sine * sine, sin(delta));
factor = 1.0; // start transform factor
for (size_t j = 0; j < i; ++j) // iterations through groups
// with different transform factors
{
for (size_t k = j; k < num_elements; k += next) // iterations through
// pairs within group
{
match = k + i;
fftTransformHelper(data, match, k, product, factor);
}
factor = mult * factor + factor;
}
}
return true;
}
// Generic template for complex FFT followed by its explicit specializations
template <class TComplexArray, int NumDims>
struct CFFT
{};
// 1D FFT:
template <class TComplexArray1D>
struct CFFT<TComplexArray1D,1>
{
// NOTE: passing by pointer is needed to avoid using element access operator
static bool FFT_inplace(TComplexArray1D & data, const size_t size,
const FFT_direction fft_direction,
const char *& error_description)
{
if(!checkNumElements(size, error_description)) {
return false;
}
rearrangeData(data, size);
if(!makeTransform(data, size, fft_direction, error_description)) {
return false;
}
if (FFT_BACKWARD == fft_direction) {
scaleValues(data, size);
}
return true;
}
};
// 2D FFT
template <class TComplexArray2D>
struct CFFT<TComplexArray2D,2>
{
static bool FFT_inplace(TComplexArray2D & data, const size_t size1, const size_t size2,
const FFT_direction fft_direction, const char *& error_description)
{
int n_rows = static_cast<int>(size1);
int n_cols = static_cast<int>(size2);
// fft for columns
std::vector<complex_type> subarray(n_rows); // each column has n_rows elements
for(int j = 0; j < n_cols; ++j)
{
#ifndef __clang__
#ifdef __USE_OPENMP
#pragma omp parallel for
#endif
#endif
for(int i = 0; i < n_rows; ++i) {
#ifdef __USE_SQUARE_BRACKETS_FOR_ELEMENT_ACCESS_OPERATOR
subarray[i] = data[i][j];
#else
subarray[i] = data(i,j);
#endif
}
if(!CFFT<std::vector<complex_type>,1>::FFT_inplace(subarray, size1,
fft_direction,
error_description))
{
return false;
}
#ifndef __clang__
#ifdef __USE_OPENMP
#pragma omp parallel for
#endif
#endif
for(int i = 0; i < n_rows; ++i) {
#ifdef __USE_SQUARE_BRACKETS_FOR_ELEMENT_ACCESS_OPERATOR
data[i][j] = subarray[i];
#else
data(i,j) = subarray[i];
#endif
}
}
// fft for rows
subarray.resize(n_cols); // each row has n_cols elements
for(int i = 0; i < n_rows; ++i)
{
#ifndef __clang__
#ifdef __USE_OPENMP
#pragma omp parallel for
#endif
#endif
for(int j = 0; j < n_cols; ++j) {
#ifdef __USE_SQUARE_BRACKETS_FOR_ELEMENT_ACCESS_OPERATOR
subarray[j] = data[i][j];
#else
subarray[j] = data(i,j);
#endif
}
if(!CFFT<std::vector<complex_type>,1>::FFT_inplace(subarray, size2,
fft_direction,
error_description))
{
return false;
}
#ifndef __clang__
#ifdef __USE_OPENMP
#pragma omp parallel for
#endif
#endif
for(int j = 0; j < n_cols; ++j) {
#ifdef __USE_SQUARE_BRACKETS_FOR_ELEMENT_ACCESS_OPERATOR
data[i][j] = subarray[j];
#else
data(i,j) = subarray[j];
#endif
}
}
return true;
}
};
// 3D FFT
template <class TComplexArray3D>
struct CFFT<TComplexArray3D,3>
{
static bool FFT_inplace(TComplexArray3D & data, const size_t size1, const size_t size2,
const size_t size3, const FFT_direction fft_direction,
const char *& error_description)
{
int n_rows = static_cast<int>(size1);
int n_cols = static_cast<int>(size2);
int n_depth = static_cast<int>(size3);
std::vector<complex_type> subarray(n_rows); // for fft for columns: each column has n_rows elements
for(int k = 0; k < n_depth; ++k) // for all depth layers
{
// fft for columns
for(int j = 0; j < n_cols; ++j)
{
#ifndef __clang__
#ifdef __USE_OPENMP
#pragma omp parallel for
#endif
#endif
for(int i = 0; i < n_rows; ++i) {
#ifdef __USE_SQUARE_BRACKETS_FOR_ELEMENT_ACCESS_OPERATOR
subarray[i] = data[i][j][k];
#else
subarray[i] = data(i,j,k);
#endif
}
if(!CFFT<std::vector<complex_type>,1>::FFT_inplace(subarray, size1,
fft_direction,
error_description))
{
return false;
}
#ifndef __clang__
#ifdef __USE_OPENMP
#pragma omp parallel for
#endif
#endif
for(int i = 0; i < n_rows; ++i) {
#ifdef __USE_SQUARE_BRACKETS_FOR_ELEMENT_ACCESS_OPERATOR
data[i][j][k] = subarray[i];
#else
data(i,j,k) = subarray[i];
#endif
}
}
}
subarray.resize(n_cols); // for fft for rows: each row has n_cols elements
for(int k = 0; k < n_depth; ++k) // for all depth layers
{
// fft for rows
for(int i = 0; i < n_rows; ++i)
{
#ifndef __clang__
#ifdef __USE_OPENMP
#pragma omp parallel for
#endif
#endif
for(int j = 0; j < n_cols; ++j) {
#ifdef __USE_SQUARE_BRACKETS_FOR_ELEMENT_ACCESS_OPERATOR
subarray[j] = data[i][j][k];
#else
subarray[j] = data(i,j,k);
#endif
}
if(!CFFT<std::vector<complex_type>,1>::FFT_inplace(subarray, size2,
fft_direction,
error_description))
{
return false;
}
#ifndef __clang__
#ifdef __USE_OPENMP
#pragma omp parallel for
#endif
#endif
for(int j = 0; j < n_cols; ++j) {
#ifdef __USE_SQUARE_BRACKETS_FOR_ELEMENT_ACCESS_OPERATOR
data[i][j][k] = subarray[j];
#else
data(i,j,k) = subarray[j];
#endif
}
}
}
// fft for depth
subarray.resize(n_depth); // each depth strip contains n_depth elements
for(int i = 0; i < n_rows; ++i) // for all rows layers
{
for(int j = 0; j < n_cols; ++j) // for all cols layers
{
#ifndef __clang__
#ifdef __USE_OPENMP
#pragma omp parallel for
#endif
#endif
for(int k = 0; k < n_depth; ++k) {
#ifdef __USE_SQUARE_BRACKETS_FOR_ELEMENT_ACCESS_OPERATOR
subarray[k] = data[i][j][k];
#else
subarray[k] = data(i,j,k);
#endif
}
if(!CFFT<std::vector<complex_type>,1>::FFT_inplace(subarray, size3,
fft_direction,
error_description))
{
return false;
}
#ifndef __clang__
#ifdef __USE_OPENMP
#pragma omp parallel for
#endif
#endif
for(int k = 0; k < n_depth; ++k) {
#ifdef __USE_SQUARE_BRACKETS_FOR_ELEMENT_ACCESS_OPERATOR
data[i][j][k] = subarray[k];
#else
data(i,j,k) = subarray[k];
#endif
}
}
}
return true;
}
};
} // namespace impl
} // namespace simple_fft
#endif // __SIMPLE_FFT__FFT_IMPL_HPP__