Page MenuHomec4science
Files F91418845

fftwmpi_engine.cc
No OneTemporary
Actions

File Metadata

Created
Sun, Nov 10, 22:32

fftwmpi_engine.cc
View Options

/**
* @file fftwmpi_engine.cc
*
* @author Lars Pastewka <lars.pastewka@imtek.uni-freiburg.de>
*
* @date 06 Mar 2017
*
* @brief implements the MPI-parallel fftw engine
*
* Copyright © 2017 Till Junge
*
* µSpectre is free software; you can redistribute it and/or
* modify it under the terms of the GNU General Public License as
* published by the Free Software Foundation, either version 3, or (at
* your option) any later version.
*
* µSpectre 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
* General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with GNU Emacs; see the file COPYING. If not, write to the
* Free Software Foundation, Inc., 59 Temple Place - Suite 330,
* Boston, MA 02111-1307, USA.
*/
#include "common/ccoord_operations.hh"
#include "fft/fftwmpi_engine.hh"
namespace muSpectre {
template <Dim_t DimsS, Dim_t DimM>
int FFTWMPIEngine<DimsS, DimM>::nb_engines{0};
template <Dim_t DimS, Dim_t DimM>
FFTWMPIEngine<DimS, DimM>::FFTWMPIEngine(Ccoord resolutions, Rcoord lengths,
Communicator comm)
:Parent{resolutions, lengths, comm}
{
if (!this->nb_engines) fftw_mpi_init();
this->nb_engines++;
std::array<ptrdiff_t, DimS> narr;
std::copy(this->domain_resolutions.begin(), this->domain_resolutions.end(),
narr.begin());
narr[DimS-1] = this->domain_resolutions[DimS-1]/2+1;
ptrdiff_t res_x, loc_x, res_y, loc_y;
this->workspace_size =
fftw_mpi_local_size_many_transposed(DimS, narr.data(),
Field_t::nb_components,
FFTW_MPI_DEFAULT_BLOCK,
FFTW_MPI_DEFAULT_BLOCK,
this->comm.get_mpi_comm(),
&res_x, &loc_x, &res_y, &loc_y);
this->fourier_resolutions[1] = this->fourier_resolutions[0];
this->fourier_locations[1] = this->fourier_locations[0];
this->resolutions[0] = res_x;
this->locations[0] = loc_x;
this->fourier_resolutions[0] = res_y;
this->fourier_locations[0] = loc_y;
for (auto & n: this->resolutions) {
if (n == 0) {
throw std::runtime_error("FFTW MPI planning returned zero resolution. "
"You may need to run on fewer processes.");
}
}
for (auto & n: this->fourier_resolutions) {
if (n == 0) {
throw std::runtime_error("FFTW MPI planning returned zero Fourier "
"resolution. You may need to run on fewer "
"processes.");
}
}
for (auto && pixel:
std::conditional_t<
DimS==2,
CcoordOps::Pixels<DimS, 1, 0>,
CcoordOps::Pixels<DimS, 1, 0, 2>
>(this->fourier_resolutions, this->fourier_locations)) {
this->work_space_container.add_pixel(pixel);
}
}
/* ---------------------------------------------------------------------- */
template <Dim_t DimS, Dim_t DimM>
void FFTWMPIEngine<DimS, DimM>::initialise(FFT_PlanFlags plan_flags) {
if (this->initialised) {
throw std::runtime_error("double initialisation, will leak memory");
}
// Initialize parent after local resolutions have been determined and
// work space has been initialized
Parent::initialise(plan_flags);
this->real_workspace = fftw_alloc_real(2*this->workspace_size);
// We need to check whether the workspace provided by our field is large
// enough. MPI parallel FFTW may request a workspace size larger than the
// nominal size of the complex buffer.
if (long(this->work.size()*Field_t::nb_components) < this->workspace_size) {
this->work.set_pad_size(this->workspace_size -
Field_t::nb_components*this->work.size());
}
unsigned int flags;
switch (plan_flags) {
case FFT_PlanFlags::estimate: {
flags = FFTW_ESTIMATE;
break;
}
case FFT_PlanFlags::measure: {
flags = FFTW_MEASURE;
break;
}
case FFT_PlanFlags::patient: {
flags = FFTW_PATIENT;
break;
}
default:
throw std::runtime_error("unknown planner flag type");
break;
}
std::array<ptrdiff_t, DimS> narr;
std::copy(this->domain_resolutions.begin(), this->domain_resolutions.end(),
narr.begin());
Real * in{this->real_workspace};
fftw_complex * out{reinterpret_cast<fftw_complex*>(this->work.data())};
this->plan_fft = fftw_mpi_plan_many_dft_r2c(
DimS, narr.data(), Field_t::nb_components, FFTW_MPI_DEFAULT_BLOCK,
FFTW_MPI_DEFAULT_BLOCK, in, out, this->comm.get_mpi_comm(),
FFTW_MPI_TRANSPOSED_OUT | flags);
if (this->plan_fft == nullptr) {
throw std::runtime_error("r2c plan failed");
}
fftw_complex * i_in = reinterpret_cast<fftw_complex*>(this->work.data());
Real * i_out = this->real_workspace;
this->plan_ifft = fftw_mpi_plan_many_dft_c2r(
DimS, narr.data(), Field_t::nb_components, FFTW_MPI_DEFAULT_BLOCK,
FFTW_MPI_DEFAULT_BLOCK, i_in, i_out, this->comm.get_mpi_comm(),
FFTW_MPI_TRANSPOSED_IN | flags);
if (this->plan_ifft == nullptr) {
throw std::runtime_error("c2r plan failed");
}
this->initialised = true;
}
/* ---------------------------------------------------------------------- */
template <Dim_t DimS, Dim_t DimM>
FFTWMPIEngine<DimS, DimM>::~FFTWMPIEngine<DimS, DimM>() noexcept {
if (this->real_workspace != nullptr) fftw_free(this->real_workspace);
if (this->plan_fft != nullptr) fftw_destroy_plan(this->plan_fft);
if (this->plan_ifft != nullptr) fftw_destroy_plan(this->plan_ifft);
// TODO: We cannot run fftw_mpi_cleanup since also calls fftw_cleanup
// and any running FFTWEngine will fail afterwards.
//this->nb_engines--;
//if (!this->nb_engines) fftw_mpi_cleanup();
}
/* ---------------------------------------------------------------------- */
template <Dim_t DimS, Dim_t DimM>
typename FFTWMPIEngine<DimS, DimM>::Workspace_t &
FFTWMPIEngine<DimS, DimM>::fft (Field_t & field) {
if (this->plan_fft == nullptr) {
throw std::runtime_error("fft plan not initialised");
}
if (field.size() != CcoordOps::get_size(this->resolutions)) {
throw std::runtime_error("size mismatch");
}
// Copy non-padded field to padded real_workspace.
// Transposed output of M x N x L transform for >= 3 dimensions is padded
// M x N x 2*(L/2+1).
ptrdiff_t fstride = Field_t::nb_components*this->resolutions[DimS-1];
ptrdiff_t wstride = Field_t::nb_components*2*(this->resolutions[DimS-1]/2+1);
ptrdiff_t n = field.size()/this->resolutions[DimS-1];
auto fdata = field.data();
auto wdata = this->real_workspace;
for (int i = 0; i < n; i++) {
std::copy(fdata, fdata+fstride, wdata);
fdata += fstride;
wdata += wstride;
}
// Compute FFT
fftw_mpi_execute_dft_r2c(
this->plan_fft, this->real_workspace,
reinterpret_cast<fftw_complex*>(this->work.data()));
return this->work;
}
/* ---------------------------------------------------------------------- */
template <Dim_t DimS, Dim_t DimM>
void
FFTWMPIEngine<DimS, DimM>::ifft (Field_t & field) const {
if (this->plan_ifft == nullptr) {
throw std::runtime_error("ifft plan not initialised");
}
if (field.size() != CcoordOps::get_size(this->resolutions)) {
throw std::runtime_error("size mismatch");
}
// Compute inverse FFT
fftw_mpi_execute_dft_c2r(
this->plan_ifft, reinterpret_cast<fftw_complex*>(this->work.data()),
this->real_workspace);
// Copy non-padded field to padded real_workspace.
// Transposed output of M x N x L transform for >= 3 dimensions is padded
// M x N x 2*(L/2+1).
ptrdiff_t fstride{Field_t::nb_components*this->resolutions[DimS-1]};
ptrdiff_t wstride{Field_t::nb_components*2*(this->resolutions[DimS-1]/2+1)};
ptrdiff_t n(field.size()/this->resolutions[DimS-1]);
auto fdata{field.data()};
auto wdata{this->real_workspace};
for (int i = 0; i < n; i++) {
std::copy(wdata, wdata+fstride, fdata);
fdata += fstride;
wdata += wstride;
}
}
template class FFTWMPIEngine<twoD, twoD>;
template class FFTWMPIEngine<twoD, threeD>;
template class FFTWMPIEngine<threeD, threeD>;
} // muSpectre

Event Timeline

c4science · Help