guilhem_lavaux/cosmotool
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Reformat and add 1d support

Browse Source
This commit is contained in:
Guilhem Lavaux 2021-04-22 08:27:18 +02:00
parent e6950440a3
commit 3e013139f2
2 changed files with 218 additions and 177 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

146
src/fourier/fft/fftw_calls.hpp
View File

@ -39,91 +39,107 @@ knowledge of the CeCILL license and that you accept its terms.
#include <fftw3.h> #include <fftw3.h>
#include <complex> #include <complex>
namespace CosmoTool namespace CosmoTool {
{
static inline void init_fftw_wisdom() static inline void init_fftw_wisdom() {
{
fftw_import_system_wisdom(); fftw_import_system_wisdom();
fftw_import_wisdom_from_filename("fft_wisdom"); fftw_import_wisdom_from_filename("fft_wisdom");
} }
static inline void save_fftw_wisdom() static inline void save_fftw_wisdom() {
{
fftw_export_wisdom_to_filename("fft_wisdom"); fftw_export_wisdom_to_filename("fft_wisdom");
} }
template<typename T> class FFTW_Calls {};
template <typename T>
class FFTW_Calls {};
#define FFTW_CALLS_BASE(rtype, prefix) \ #define FFTW_CALLS_BASE(rtype, prefix) \
template<> \ template <> \
class FFTW_Calls<rtype> { \ class FFTW_Calls<rtype> { \
public: \ public: \
typedef rtype real_type; \ typedef rtype real_type; \
typedef prefix ## _complex complex_type; \ typedef prefix##_complex complex_type; \
typedef prefix ## _plan plan_type; \ typedef prefix##_plan plan_type; \
\ \
static complex_type *alloc_complex(size_t N) { return prefix ## _alloc_complex(N); } \ static complex_type *alloc_complex(size_t N) { \
static real_type *alloc_real(size_t N) { return prefix ## _alloc_real(N); } \ return prefix##_alloc_complex(N); \
} \
static real_type *alloc_real(size_t N) { return prefix##_alloc_real(N); } \
static void free(void *p) { fftw_free(p); } \ static void free(void *p) { fftw_free(p); } \
\ \
static void execute(plan_type p) { prefix ## _execute(p); } \ static void execute(plan_type p) { prefix##_execute(p); } \
static void execute_r2c(plan_type p, real_type *in, complex_type *out) { prefix ## _execute_dft_r2c(p, in, out); } \ static void execute_r2c(plan_type p, real_type *in, complex_type *out) { \
static void execute_c2r(plan_type p, complex_type *in, real_type *out) { prefix ## _execute_dft_c2r(p, in, out); } \ prefix##_execute_dft_r2c(p, in, out); \
static void execute_r2c(plan_type p, real_type *in, std::complex<real_type> *out) { prefix ## _execute_dft_r2c(p, in, (complex_type*)out); } \
static void execute_c2r(plan_type p, std::complex<real_type> *in, real_type *out) { prefix ## _execute_dft_c2r(p, (complex_type*) in, out); } \
static void execute_c2c(plan_type p, std::complex<real_type> *in, std::complex<real_type> *out) { prefix ## _execute_dft(p, (complex_type *)in, (complex_type*)out); } \
static plan_type plan_dft_r2c_2d(int Nx, int Ny, \
real_type *in, complex_type *out, \
unsigned flags) \
{ \
return prefix ## _plan_dft_r2c_2d(Nx, Ny, in, out, \
flags); \
} \ } \
static plan_type plan_dft_c2r_2d(int Nx, int Ny, \ static void execute_c2r(plan_type p, complex_type *in, real_type *out) { \
complex_type *in, real_type *out, \ prefix##_execute_dft_c2r(p, in, out); \
unsigned flags) \
{ \
return prefix ## _plan_dft_c2r_2d(Nx, Ny, in, out, \
flags); \
} \ } \
static plan_type plan_dft_r2c_3d(int Nx, int Ny, int Nz, \ static void \
real_type *in, complex_type *out, \ execute_r2c(plan_type p, real_type *in, std::complex<real_type> *out) { \
unsigned flags) \ prefix##_execute_dft_r2c(p, in, (complex_type *)out); \
{ \
return prefix ## _plan_dft_r2c_3d(Nx, Ny, Nz, in, out, flags); \
} \ } \
static plan_type plan_dft_c2r_3d(int Nx, int Ny, int Nz, \ static void \
complex_type *in, real_type *out, \ execute_c2r(plan_type p, std::complex<real_type> *in, real_type *out) { \
unsigned flags) \ prefix##_execute_dft_c2r(p, (complex_type *)in, out); \
{ \
return prefix ## _plan_dft_c2r_3d(Nx, Ny, Nz, in, out, flags); \
} \ } \
\ static void execute_c2c( \
static plan_type plan_dft_r2c(int rank, const int *n, real_type *in, \ plan_type p, std::complex<real_type> *in, \
complex_type *out, unsigned flags) \ std::complex<real_type> *out) { \
{ \ prefix##_execute_dft(p, (complex_type *)in, (complex_type *)out); \
return prefix ## _plan_dft_r2c(rank, n, in, out, flags); \
} \ } \
static plan_type plan_dft_c2r(int rank, const int *n, complex_type *in, \ static plan_type plan_dft_r2c_1d( \
real_type *out, unsigned flags) \ int Nx, real_type *in, complex_type *out, unsigned flags) { \
{ \ return prefix##_plan_dft_r2c_1d(Nx, in, out, flags); \
return prefix ## _plan_dft_c2r(rank, n, in, out, flags); \
} \ } \
static plan_type plan_dft_3d(int Nx, int Ny, int Nz, complex_type *in, complex_type *out, int sign, unsigned flags) { \ static plan_type plan_dft_c2r_1d( \
return prefix ## _plan_dft_3d(Nx, Ny, Nz, in, out, sign, flags); \ int Nx, complex_type *in, real_type *out, unsigned flags) { \
return prefix##_plan_dft_c2r_1d(Nx, in, out, flags); \
} \ } \
static plan_type plan_dft_2d(int Nx, int Ny, complex_type *in, complex_type *out, int sign, unsigned flags) { \ static plan_type plan_dft_r2c_2d( \
return prefix ## _plan_dft_2d(Nx, Ny, in, out, sign, flags); \ int Nx, int Ny, real_type *in, complex_type *out, unsigned flags) { \
return prefix##_plan_dft_r2c_2d(Nx, Ny, in, out, flags); \
} \ } \
static void destroy_plan(plan_type plan) { prefix ## _destroy_plan(plan); } \ static plan_type plan_dft_c2r_2d( \
} int Nx, int Ny, complex_type *in, real_type *out, unsigned flags) { \
return prefix##_plan_dft_c2r_2d(Nx, Ny, in, out, flags); \
} \
static plan_type plan_dft_r2c_3d( \
int Nx, int Ny, int Nz, real_type *in, complex_type *out, \
unsigned flags) { \
return prefix##_plan_dft_r2c_3d(Nx, Ny, Nz, in, out, flags); \
} \
static plan_type plan_dft_c2r_3d( \
int Nx, int Ny, int Nz, complex_type *in, real_type *out, \
unsigned flags) { \
return prefix##_plan_dft_c2r_3d(Nx, Ny, Nz, in, out, flags); \
} \
\
static plan_type plan_dft_r2c( \
int rank, const int *n, real_type *in, complex_type *out, \
unsigned flags) { \
return prefix##_plan_dft_r2c(rank, n, in, out, flags); \
} \
static plan_type plan_dft_c2r( \
int rank, const int *n, complex_type *in, real_type *out, \
unsigned flags) { \
return prefix##_plan_dft_c2r(rank, n, in, out, flags); \
} \
static plan_type plan_dft_3d( \
int Nx, int Ny, int Nz, complex_type *in, complex_type *out, int sign, \
unsigned flags) { \
return prefix##_plan_dft_3d(Nx, Ny, Nz, in, out, sign, flags); \
} \
static plan_type plan_dft_2d( \
int Nx, int Ny, complex_type *in, complex_type *out, int sign, \
unsigned flags) { \
return prefix##_plan_dft_2d(Nx, Ny, in, out, sign, flags); \
} \
static void destroy_plan(plan_type plan) { prefix##_destroy_plan(plan); } \
}
FFTW_CALLS_BASE(double, fftw);
FFTW_CALLS_BASE(double, fftw); FFTW_CALLS_BASE(float, fftwf);
FFTW_CALLS_BASE(float, fftwf);
#undef FFTW_CALLS_BASE #undef FFTW_CALLS_BASE
}; }; // namespace CosmoTool
#endif #endif

185
src/fourier/fft/fftw_calls_mpi.hpp
View File

@ -5,110 +5,135 @@
#include <mpi.h> #include <mpi.h>
#include <fftw3-mpi.h> #include <fftw3-mpi.h>
namespace CosmoTool namespace CosmoTool {
{
static inline void init_fftw_mpi() static inline void init_fftw_mpi() { fftw_mpi_init(); }
{
fftw_mpi_init();
}
static inline void done_fftw_mpi() static inline void done_fftw_mpi() { fftw_mpi_cleanup(); }
{
fftw_mpi_cleanup();
}
template<typename T> class FFTW_MPI_Calls {};
template <typename T>
class FFTW_MPI_Calls {};
#define FFTW_MPI_CALLS_BASE(rtype, prefix) \ #define FFTW_MPI_CALLS_BASE(rtype, prefix) \
template<> \ template <> \
class FFTW_MPI_Calls<rtype> { \ class FFTW_MPI_Calls<rtype> { \
public: \ public: \
typedef rtype real_type; \ typedef rtype real_type; \
typedef prefix ## _complex complex_type; \ typedef prefix##_complex complex_type; \
typedef prefix ## _plan plan_type; \ typedef prefix##_plan plan_type; \
\ \
static complex_type *alloc_complex(size_t N) { return prefix ## _alloc_complex(N); } \ static complex_type *alloc_complex(size_t N) { \
static real_type *alloc_real(size_t N) { return prefix ## _alloc_real(N); } \ return prefix##_alloc_complex(N); \
} \
static real_type *alloc_real(size_t N) { return prefix##_alloc_real(N); } \
static void free(void *p) { fftw_free(p); } \ static void free(void *p) { fftw_free(p); } \
\ \
template<size_t Nd> \ template <size_t Nd> \
static ptrdiff_t local_size(std::array<ptrdiff_t,Nd> const& N, MPI_Comm comm, \ static ptrdiff_t local_size( \
std::array<ptrdiff_t, Nd> const &N, MPI_Comm comm, \
ptrdiff_t *local_n0, ptrdiff_t *local_0_start) { \ ptrdiff_t *local_n0, ptrdiff_t *local_0_start) { \
return prefix ## _mpi_local_size(Nd, N.data(), comm, local_n0, local_0_start); \ return prefix##_mpi_local_size( \
Nd, N.data(), comm, local_n0, local_0_start); \
} \ } \
static ptrdiff_t local_size_2d(ptrdiff_t N0, ptrdiff_t N1, MPI_Comm comm, \ static ptrdiff_t local_size_2d( \
ptrdiff_t N0, ptrdiff_t N1, MPI_Comm comm, ptrdiff_t *local_n0, \
ptrdiff_t *local_0_start) { \
return prefix##_mpi_local_size_2d( \
N0, N1, comm, local_n0, local_0_start); \
} \
\
static ptrdiff_t local_size_3d( \
ptrdiff_t N0, ptrdiff_t N1, ptrdiff_t N2, MPI_Comm comm, \
ptrdiff_t *local_n0, ptrdiff_t *local_0_start) { \ ptrdiff_t *local_n0, ptrdiff_t *local_0_start) { \
return prefix ## _mpi_local_size_2d(N0, N1, comm, local_n0, local_0_start); \ return prefix##_mpi_local_size_3d( \
N0, N1, N2, comm, local_n0, local_0_start); \
} \ } \
\ \
static ptrdiff_t local_size_3d(ptrdiff_t N0, ptrdiff_t N1, ptrdiff_t N2, MPI_Comm comm, \ static void execute(plan_type p) { prefix##_execute(p); } \
ptrdiff_t *local_n0, ptrdiff_t *local_0_start) { \ static void \
return prefix ## _mpi_local_size_3d(N0, N1, N2, comm, local_n0, local_0_start); \ execute_c2c(plan_type p, complex_type *in, complex_type *out) { \
prefix##_mpi_execute_dft(p, in, out); \
} \ } \
\ static void execute_c2c( \
static void execute(plan_type p) { prefix ## _execute(p); } \ plan_type p, std::complex<real_type> *in, \
static void execute_c2c(plan_type p, complex_type *in, complex_type *out) { prefix ## _mpi_execute_dft(p, in, out); } \ std::complex<real_type> *out) { \
static void execute_c2c(plan_type p, std::complex<real_type> *in, std::complex<real_type> *out) { prefix ## _mpi_execute_dft(p, (complex_type*)in, (complex_type*)out); } \ prefix##_mpi_execute_dft(p, (complex_type *)in, (complex_type *)out); \
static void execute_r2c(plan_type p, real_type *in, complex_type *out) { prefix ## _mpi_execute_dft_r2c(p, in, out); } \
static void execute_c2r(plan_type p, std::complex<real_type> *in, real_type *out) { prefix ## _mpi_execute_dft_c2r(p, (complex_type*)in, out); } \
static void execute_c2r(plan_type p, complex_type *in, real_type *out) { prefix ## _mpi_execute_dft_c2r(p, in, out); } \
static void execute_r2c(plan_type p, real_type *in, std::complex<real_type> *out) { prefix ## _mpi_execute_dft_r2c(p, in, (complex_type*)out); } \
\
static plan_type plan_dft_r2c_2d(int Nx, int Ny, \
real_type *in, complex_type *out, \
MPI_Comm comm, unsigned flags) \
{ \
return prefix ## _mpi_plan_dft_r2c_2d(Nx, Ny, in, out, \
comm, flags); \
} \ } \
static plan_type plan_dft_c2r_2d(int Nx, int Ny, \ static void execute_r2c(plan_type p, real_type *in, complex_type *out) { \
complex_type *in, real_type *out, \ prefix##_mpi_execute_dft_r2c(p, in, out); \
MPI_Comm comm, unsigned flags) \
{ \
return prefix ## _mpi_plan_dft_c2r_2d(Nx, Ny, in, out, \
comm, flags); \
} \ } \
static plan_type plan_dft_r2c_3d(int Nx, int Ny, int Nz, \ static void \
real_type *in, complex_type *out, \ execute_c2r(plan_type p, std::complex<real_type> *in, real_type *out) { \
MPI_Comm comm, unsigned flags) \ prefix##_mpi_execute_dft_c2r(p, (complex_type *)in, out); \
{ \
return prefix ## _mpi_plan_dft_r2c_3d(Nx, Ny, Nz, in, out, comm, flags); \
} \ } \
static plan_type plan_dft_c2r_3d(int Nx, int Ny, int Nz, \ static void execute_c2r(plan_type p, complex_type *in, real_type *out) { \
complex_type *in, real_type *out, \ prefix##_mpi_execute_dft_c2r(p, in, out); \
MPI_Comm comm, \
unsigned flags) \
{ \
return prefix ## _mpi_plan_dft_c2r_3d(Nx, Ny, Nz, in, out, comm, flags); \
} \ } \
\ static void \
static plan_type plan_dft_r2c(int rank, const ptrdiff_t *n, real_type *in, \ execute_r2c(plan_type p, real_type *in, std::complex<real_type> *out) { \
complex_type *out, MPI_Comm comm, unsigned flags) \ prefix##_mpi_execute_dft_r2c(p, in, (complex_type *)out); \
{ \
return prefix ## _mpi_plan_dft_r2c(rank, n, in, out, comm, flags); \
} \ } \
static plan_type plan_dft_c2r(int rank, const ptrdiff_t *n, complex_type *in, \ \
real_type *out, MPI_Comm comm, unsigned flags) \ static plan_type plan_dft_r2c_1d( \
{ \ int n, real_type *in, complex_type *out, MPI_Comm, unsigned flags) { \
return prefix ## _mpi_plan_dft_c2r(rank, n, in, out, comm, flags); \ return prefix##_plan_dft_r2c_1d(n, in, out, flags); \
} \ } \
static plan_type plan_dft_3d(int Nx, int Ny, int Nz, complex_type *in, complex_type *out, MPI_Comm comm, int sign, unsigned flags) { \ \
return prefix ## _mpi_plan_dft_3d(Nx, Ny, Nz, in, out, comm, sign, flags); \ static plan_type plan_dft_r2c_2d( \
int Nx, int Ny, real_type *in, complex_type *out, MPI_Comm comm, \
unsigned flags) { \
return prefix##_mpi_plan_dft_r2c_2d(Nx, Ny, in, out, comm, flags); \
} \ } \
static plan_type plan_dft_2d(int Nx, int Ny, complex_type *in, complex_type *out, MPI_Comm comm, int sign, unsigned flags) { \ \
return prefix ## _mpi_plan_dft_2d(Nx, Ny, in, out, comm, sign, flags); \ static plan_type plan_dft_r2c_1d( \
int n, complex_type *in, real_type *out, MPI_Comm, unsigned flags) { \
return prefix##_plan_dft_c2r_1d(n, in, out, flags); \
} \ } \
static void destroy_plan(plan_type plan) { prefix ## _destroy_plan(plan); } \ static plan_type plan_dft_c2r_2d( \
} int Nx, int Ny, complex_type *in, real_type *out, MPI_Comm comm, \
unsigned flags) { \
return prefix##_mpi_plan_dft_c2r_2d(Nx, Ny, in, out, comm, flags); \
} \
\
static plan_type plan_dft_r2c_3d( \
int Nx, int Ny, int Nz, real_type *in, complex_type *out, \
MPI_Comm comm, unsigned flags) { \
return prefix##_mpi_plan_dft_r2c_3d(Nx, Ny, Nz, in, out, comm, flags); \
} \
static plan_type plan_dft_c2r_3d( \
int Nx, int Ny, int Nz, complex_type *in, real_type *out, \
MPI_Comm comm, unsigned flags) { \
return prefix##_mpi_plan_dft_c2r_3d(Nx, Ny, Nz, in, out, comm, flags); \
} \
\
static plan_type plan_dft_r2c( \
int rank, const ptrdiff_t *n, real_type *in, complex_type *out, \
MPI_Comm comm, unsigned flags) { \
return prefix##_mpi_plan_dft_r2c(rank, n, in, out, comm, flags); \
} \
static plan_type plan_dft_c2r( \
int rank, const ptrdiff_t *n, complex_type *in, real_type *out, \
MPI_Comm comm, unsigned flags) { \
return prefix##_mpi_plan_dft_c2r(rank, n, in, out, comm, flags); \
} \
static plan_type plan_dft_3d( \
int Nx, int Ny, int Nz, complex_type *in, complex_type *out, \
MPI_Comm comm, int sign, unsigned flags) { \
return prefix##_mpi_plan_dft_3d(Nx, Ny, Nz, in, out, comm, sign, flags); \
} \
static plan_type plan_dft_2d( \
int Nx, int Ny, complex_type *in, complex_type *out, MPI_Comm comm, \
int sign, unsigned flags) { \
return prefix##_mpi_plan_dft_2d(Nx, Ny, in, out, comm, sign, flags); \
} \
static void destroy_plan(plan_type plan) { prefix##_destroy_plan(plan); } \
}
FFTW_MPI_CALLS_BASE(double, fftw);
FFTW_MPI_CALLS_BASE(double, fftw); FFTW_MPI_CALLS_BASE(float, fftwf);
FFTW_MPI_CALLS_BASE(float, fftwf);
#undef FFTW_MPI_CALLS_BASE #undef FFTW_MPI_CALLS_BASE
}; }; // namespace CosmoTool
#endif #endif