Convolve.cpp

Go to the documentation of this file.

//  ************************************************************************************************
//
//  BornAgain: simulate and fit reflection and scattering
//
//! @file      Device/Resolution/Convolve.cpp
//! @brief     Implements class Convolve.
//!
//! @homepage  http://www.bornagainproject.org
//! @license   GNU General Public License v3 or higher (see COPYING)
//! @copyright Forschungszentrum Jülich GmbH 2018
//! @authors   Scientific Computing Group at MLZ (see CITATION, AUTHORS)
//
//  ************************************************************************************************
 
#include "Device/Resolution/Convolve.h"
#include <iostream>
#include <sstream>
#include <stdexcept> // need overlooked by g++ 5.4
 
Convolve::Convolve()
    : m_mode(FFTW_UNDEFINED)
{
    // storing favorite fftw3 prime factors
    const size_t FFTW_FACTORS[] = {13, 11, 7, 5, 3, 2};
    m_implemented_factors.assign(FFTW_FACTORS,
                                 FFTW_FACTORS + sizeof(FFTW_FACTORS) / sizeof(FFTW_FACTORS[0]));
}
 
Convolve::Workspace::~Workspace()
{
    clear();
}
 
void Convolve::Workspace::clear()
{
    h_src = 0;
    w_src = 0;
    h_kernel = 0;
    w_kernel = 0;
 
    delete[] in_src;
    in_src = nullptr;
 
    if (out_src)
        fftw_free((fftw_complex*)out_src);
    out_src = nullptr;
 
 
    delete[] in_kernel;
    in_kernel = nullptr;
 
    if (out_kernel)
        fftw_free((fftw_complex*)out_kernel);
    out_kernel = nullptr;
 
 
    delete[] dst_fft;
    dst_fft = nullptr;
 
    // if(dst) delete[] dst;
    // dst=0;
 
    h_offset = 0;
    w_offset = 0;
 
    if (p_forw_src != nullptr)
        fftw_destroy_plan(p_forw_src);
    if (p_forw_kernel != nullptr)
        fftw_destroy_plan(p_forw_kernel);
    if (p_back != nullptr)
        fftw_destroy_plan(p_back);
 
    fftw_cleanup();
}
 
/* ************************************************************************* */
// convolution in 2d
/* ************************************************************************* */
void Convolve::fftconvolve(const double2d_t& source, const double2d_t& kernel, double2d_t& result)
{
    // set default convolution mode, if not defined
    if (m_mode == FFTW_UNDEFINED)
        setMode(FFTW_LINEAR_SAME);
 
    int h_src = (int)source.size();
    int w_src = (int)(!source.empty() ? source[0].size() : 0);
    int h_kernel = (int)kernel.size();
    int w_kernel = (!kernel.empty() ? (int)kernel[0].size() : 0);
 
    // initialization
    init(h_src, w_src, h_kernel, w_kernel);
    // Compute the circular convolution
    fftw_circular_convolution(source, kernel);
 
    // results
    result.clear();
    result.resize(ws.h_dst);
    for (int i = 0; i < ws.h_dst; i++) {
        result[i].resize(ws.w_dst, 0);
        for (int j = 0; j < ws.w_dst; j++) {
            // result[i][j]=ws.dst[i*ws.w_dst+j];
            if (m_mode == FFTW_CIRCULAR_SAME_SHIFTED) {
                result[i][j] = ws.dst_fft[((i + int(ws.h_kernel / 2.0)) % ws.h_fftw) * ws.w_fftw
                                          + (j + int(ws.w_kernel / 2.0)) % ws.w_fftw];
            } else {
                result[i][j] = ws.dst_fft[(i + ws.h_offset) * ws.w_fftw + j + ws.w_offset];
            }
        }
    }
}
 
/* ************************************************************************* */
// convolution in 1d
/* ************************************************************************* */
void Convolve::fftconvolve(const double1d_t& source, const double1d_t& kernel, double1d_t& result)
{
    // we simply create 2d arrays with length of first dimension equal to 1, and call 2d convolution
    double2d_t source2d, kernel2d;
    source2d.push_back(source);
    kernel2d.push_back(kernel);
 
    double2d_t result2d;
    fftconvolve(source2d, kernel2d, result2d);
    if (result2d.size() != 1)
        throw std::runtime_error("Convolve::fftconvolve -> Panic in 1d");
    result = result2d[0];
}
 
/* ************************************************************************* */
// initialise input and output arrays for fast Fourier transformation
/* ************************************************************************* */
void Convolve::init(int h_src, int w_src, int h_kernel, int w_kernel)
{
    if (!h_src || !w_src || !h_kernel || !w_kernel) {
        std::ostringstream os;
        os << "Convolve::init() -> Panic! Wrong dimensions " << h_src << " " << w_src << " "
           << h_kernel << " " << w_kernel << std::endl;
        throw std::runtime_error(os.str());
    }
 
    ws.clear();
    ws.h_src = h_src;
    ws.w_src = w_src;
    ws.h_kernel = h_kernel;
    ws.w_kernel = w_kernel;
    switch (m_mode) {
    case FFTW_LINEAR_FULL:
        // Full Linear convolution
        ws.h_fftw = find_closest_factor(h_src + h_kernel - 1);
        ws.w_fftw = find_closest_factor(w_src + w_kernel - 1);
        ws.h_dst = h_src + h_kernel - 1;
        ws.w_dst = w_src + w_kernel - 1;
        // Here we just keep the first [0:h_dst-1 ; 0:w_dst-1] real part elements of out_src
        ws.h_offset = 0;
        ws.w_offset = 0;
        break;
    case FFTW_LINEAR_SAME_UNPADDED:
        // Same Linear convolution
        ws.h_fftw = h_src + int(h_kernel / 2.0);
        ws.w_fftw = w_src + int(w_kernel / 2.0);
        ws.h_dst = h_src;
        ws.w_dst = w_src;
        // Here we just keep the first [h_filt/2:h_filt/2+h_dst-1 ; w_filt/2:w_filt/2+w_dst-1]
        // real part elements of out_src
        ws.h_offset = int(ws.h_kernel / 2.0);
        ws.w_offset = int(ws.w_kernel / 2.0);
        break;
    case FFTW_LINEAR_SAME:
        // Same Linear convolution
        ws.h_fftw = find_closest_factor(h_src + int(h_kernel / 2.0));
        ws.w_fftw = find_closest_factor(w_src + int(w_kernel / 2.0));
        ws.h_dst = h_src;
        ws.w_dst = w_src;
        // Here we just keep the first [h_filt/2:h_filt/2+h_dst-1 ; w_filt/2:w_filt/2+w_dst-1]
        // real part elements of out_src
        ws.h_offset = int(ws.h_kernel / 2.0);
        ws.w_offset = int(ws.w_kernel / 2.0);
        break;
    case FFTW_LINEAR_VALID:
        // Valid Linear convolution
        if (ws.h_kernel > ws.h_src || ws.w_kernel > ws.w_src) {
            ws.h_fftw = 0;
            ws.w_fftw = 0;
            ws.h_dst = 0;
            ws.w_dst = 0;
            std::cout << "The 'valid' convolution results in an empty matrix" << std::endl;
            throw std::runtime_error("The 'valid' convolution results in an empty matrix");
        } else {
            ws.h_fftw = find_closest_factor(h_src);
            ws.w_fftw = find_closest_factor(w_src);
            ws.h_dst = h_src - h_kernel + 1;
            ws.w_dst = w_src - w_kernel + 1;
        }
        // Here we just take [h_dst x w_dst] elements starting at [h_kernel-1;w_kernel-1]
        ws.h_offset = ws.h_kernel - 1;
        ws.w_offset = ws.w_kernel - 1;
        break;
    case FFTW_CIRCULAR_SAME:
    case FFTW_CIRCULAR_SAME_SHIFTED:
        // Circular convolution, modulo N
        // We don't padd with zeros because if we do, we need to padd with at least h_kernel/2;
        // w_kernel/2 elements plus the wrapp around, which in facts leads to too much
        // computations compared to the gain obtained with the optimal size
        ws.h_fftw = h_src;
        ws.w_fftw = w_src;
        ws.h_dst = h_src;
        ws.w_dst = w_src;
        // We copy the first [0:h_dst-1 ; 0:w_dst-1] real part elements of out_src
        ws.h_offset = 0;
        ws.w_offset = 0;
        break;
    default:
        std::cout
            << "Unrecognized convolution mode, possible modes are "
            << "FFTW_LINEAR_FULL, FFTW_LINEAR_SAME, FFTW_LINEAR_SAME_UNPADDED, FFTW_LINEAR_VALID, "
            << "FFTW_CIRCULAR_SAME, FFTW_CIRCULAR_SHIFTED " << std::endl;
        break;
    }
 
    ws.in_src = new double[ws.h_fftw * ws.w_fftw];
    ws.out_src = (double*)fftw_malloc(sizeof(fftw_complex) * ws.h_fftw * (ws.w_fftw / 2 + 1));
    ws.in_kernel = new double[ws.h_fftw * ws.w_fftw];
    ws.out_kernel = (double*)fftw_malloc(sizeof(fftw_complex) * ws.h_fftw * (ws.w_fftw / 2 + 1));
 
    ws.dst_fft = new double[ws.h_fftw * ws.w_fftw];
    // ws.dst = new double[ws.h_dst * ws.w_dst];
 
    // Initialization of the plans
    ws.p_forw_src = fftw_plan_dft_r2c_2d(ws.h_fftw, ws.w_fftw, ws.in_src, (fftw_complex*)ws.out_src,
                                         FFTW_ESTIMATE);
    if (ws.p_forw_src == nullptr)
        throw std::runtime_error("Convolve::init() -> Error! Can't initialise p_forw_src plan.");
 
    ws.p_forw_kernel = fftw_plan_dft_r2c_2d(ws.h_fftw, ws.w_fftw, ws.in_kernel,
                                            (fftw_complex*)ws.out_kernel, FFTW_ESTIMATE);
    if (ws.p_forw_kernel == nullptr)
        throw std::runtime_error("Convolve::init() -> Error! Can't initialise p_forw_kernel plan.");
 
    // The backward FFT takes ws.out_kernel as input
    ws.p_back = fftw_plan_dft_c2r_2d(ws.h_fftw, ws.w_fftw, (fftw_complex*)ws.out_kernel, ws.dst_fft,
                                     FFTW_ESTIMATE);
    if (ws.p_back == nullptr)
        throw std::runtime_error("Convolve::init() -> Error! Can't initialise p_back plan.");
}
 
/* ************************************************************************* */
// initialise input and output arrays for fast Fourier transformation
/* ************************************************************************* */
 
void Convolve::fftw_circular_convolution(const double2d_t& src, const double2d_t& kernel)
{
    if (ws.h_fftw <= 0 || ws.w_fftw <= 0)
        throw std::runtime_error("Convolve::fftw_convolve() -> Panic! Initialization is missed.");
 
    double *ptr, *ptr_end, *ptr2;
 
    // Reset the content of ws.in_src
    for (ptr = ws.in_src, ptr_end = ws.in_src + ws.h_fftw * ws.w_fftw; ptr != ptr_end; ++ptr)
        *ptr = 0.0;
    for (ptr = ws.in_kernel, ptr_end = ws.in_kernel + ws.h_fftw * ws.w_fftw; ptr != ptr_end; ++ptr)
        *ptr = 0.0;
 
    // Then we build our periodic signals
    // ptr = src;
    for (int i = 0; i < ws.h_src; ++i)
        for (int j = 0; j < ws.w_src; ++j, ++ptr)
            ws.in_src[(i % ws.h_fftw) * ws.w_fftw + (j % ws.w_fftw)] += src[i][j];
 
    // ptr = kernel;
    for (int i = 0; i < ws.h_kernel; ++i)
        for (int j = 0; j < ws.w_kernel; ++j, ++ptr)
            ws.in_kernel[(i % ws.h_fftw) * ws.w_fftw + (j % ws.w_fftw)] += kernel[i][j];
 
    // And we compute their packed FFT
    fftw_execute(ws.p_forw_src);
    fftw_execute(ws.p_forw_kernel);
 
    // Compute the element-wise product on the packed terms
    // Let's put the elementwise products in ws.in_kernel
    for (ptr = ws.out_src, ptr2 = ws.out_kernel,
        ptr_end = ws.out_src + 2 * ws.h_fftw * (ws.w_fftw / 2 + 1);
         ptr != ptr_end; ++ptr, ++ptr2) {
        double re_s = *ptr;
        double im_s = *(++ptr);
        double re_k = *ptr2;
        double im_k = *(++ptr2);
        *(ptr2 - 1) = re_s * re_k - im_s * im_k;
        *ptr2 = re_s * im_k + im_s * re_k;
    }
 
    // Compute the backward FFT
    // Carefull, The backward FFT does not preserve the output
    fftw_execute(ws.p_back);
    // Scale the transform
    for (ptr = ws.dst_fft, ptr_end = ws.dst_fft + ws.w_fftw * ws.h_fftw; ptr != ptr_end; ++ptr)
        *ptr /= double(ws.h_fftw * ws.w_fftw);
}
 
/* ************************************************************************* */
// find a number closest to the given one, which  can be factorized according
// to fftw3 favorite factorization
/* ************************************************************************* */
 
int Convolve::find_closest_factor(int n)
{
    if (is_optimal(n))
        return n;
    int j = n + 1;
    while (!is_optimal(j))
        ++j;
    return j;
}
 
/* ************************************************************************* */
// if a number can be factorized using only favorite fftw3 factors
/* ************************************************************************* */
bool Convolve::is_optimal(int n)
{
    if (n == 1)
        return false;
    size_t ntest = n;
    for (size_t i = 0; i < m_implemented_factors.size(); i++) {
        while ((ntest % m_implemented_factors[i]) == 0)
            ntest = ntest / m_implemented_factors[i];
    }
    return ntest == 1;
}