cannam@95: /*
cannam@95:  * Copyright (c) 2003, 2007-11 Matteo Frigo
cannam@95:  * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
cannam@95:  *
cannam@95:  * This program is free software; you can redistribute it and/or modify
cannam@95:  * it under the terms of the GNU General Public License as published by
cannam@95:  * the Free Software Foundation; either version 2 of the License, or
cannam@95:  * (at your option) any later version.
cannam@95:  *
cannam@95:  * This program is distributed in the hope that it will be useful,
cannam@95:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
cannam@95:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
cannam@95:  * GNU General Public License for more details.
cannam@95:  *
cannam@95:  * You should have received a copy of the GNU General Public License
cannam@95:  * along with this program; if not, write to the Free Software
cannam@95:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
cannam@95:  *
cannam@95:  */
cannam@95: 
cannam@95: /* FFTW-MPI internal header file */
cannam@95: #ifndef __IFFTW_MPI_H__
cannam@95: #define __IFFTW_MPI_H__
cannam@95: 
cannam@95: #include "ifftw.h"
cannam@95: #include "rdft.h"
cannam@95: 
cannam@95: #include <mpi.h>
cannam@95: 
cannam@95: /* mpi problem flags: problem-dependent meaning, but in general
cannam@95:    SCRAMBLED means some reordering *within* the dimensions, while
cannam@95:    TRANSPOSED means some reordering *of* the dimensions */
cannam@95: #define SCRAMBLED_IN (1 << 0)
cannam@95: #define SCRAMBLED_OUT (1 << 1)
cannam@95: #define TRANSPOSED_IN (1 << 2)
cannam@95: #define TRANSPOSED_OUT (1 << 3)
cannam@95: #define RANK1_BIGVEC_ONLY (1 << 4) /* for rank=1, allow only bigvec solver */
cannam@95: 
cannam@95: #define ONLY_SCRAMBLEDP(flags) (!((flags) & ~(SCRAMBLED_IN|SCRAMBLED_OUT)))
cannam@95: #define ONLY_TRANSPOSEDP(flags) (!((flags) & ~(TRANSPOSED_IN|TRANSPOSED_OUT)))
cannam@95: 
cannam@95: #if defined(FFTW_SINGLE)
cannam@95: #  define FFTW_MPI_TYPE MPI_FLOAT
cannam@95: #elif defined(FFTW_LDOUBLE)
cannam@95: #  define FFTW_MPI_TYPE MPI_LONG_DOUBLE
cannam@95: #elif defined(FFTW_QUAD)
cannam@95: #  error MPI quad-precision type is unknown
cannam@95: #else
cannam@95: #  define FFTW_MPI_TYPE MPI_DOUBLE
cannam@95: #endif
cannam@95: 
cannam@95: /* all fftw-mpi identifiers start with fftw_mpi (or fftwf_mpi etc.) */
cannam@95: #define XM(name) X(CONCAT(mpi_, name))
cannam@95: 
cannam@95: /***********************************************************************/
cannam@95: /* block distributions */
cannam@95: 
cannam@95: /* a distributed dimension of length n with input and output block
cannam@95:    sizes ib and ob, respectively. */
cannam@95: typedef enum { IB = 0, OB } block_kind;
cannam@95: typedef struct {
cannam@95:      INT n;
cannam@95:      INT b[2]; /* b[IB], b[OB] */
cannam@95: } ddim;
cannam@95: 
cannam@95: /* Loop over k in {IB, OB}.  Note: need explicit casts for C++. */
cannam@95: #define FORALL_BLOCK_KIND(k) for (k = IB; k <= OB; k = (block_kind) (((int) k) + 1))
cannam@95: 
cannam@95: /* unlike tensors in the serial FFTW, the ordering of the dtensor
cannam@95:    dimensions matters - both the array and the block layout are
cannam@95:    row-major order. */
cannam@95: typedef struct {
cannam@95:      int rnk;
cannam@95: #if defined(STRUCT_HACK_KR)
cannam@95:      ddim dims[1];
cannam@95: #elif defined(STRUCT_HACK_C99)
cannam@95:      ddim dims[];
cannam@95: #else
cannam@95:      ddim *dims;
cannam@95: #endif
cannam@95: } dtensor;
cannam@95: 
cannam@95: 
cannam@95: /* dtensor.c: */
cannam@95: dtensor *XM(mkdtensor)(int rnk);
cannam@95: void XM(dtensor_destroy)(dtensor *sz);
cannam@95: dtensor *XM(dtensor_copy)(const dtensor *sz);
cannam@95: dtensor *XM(dtensor_canonical)(const dtensor *sz, int compress);
cannam@95: int XM(dtensor_validp)(const dtensor *sz);
cannam@95: void XM(dtensor_md5)(md5 *p, const dtensor *t);
cannam@95: void XM(dtensor_print)(const dtensor *t, printer *p);
cannam@95: 
cannam@95: /* block.c: */
cannam@95: 
cannam@95: /* for a single distributed dimension: */
cannam@95: INT XM(num_blocks)(INT n, INT block);
cannam@95: int XM(num_blocks_ok)(INT n, INT block, MPI_Comm comm);
cannam@95: INT XM(default_block)(INT n, int n_pes);
cannam@95: INT XM(block)(INT n, INT block, int which_block);
cannam@95: 
cannam@95: /* for multiple distributed dimensions: */
cannam@95: INT XM(num_blocks_total)(const dtensor *sz, block_kind k);
cannam@95: int XM(idle_process)(const dtensor *sz, block_kind k, int which_pe);
cannam@95: void XM(block_coords)(const dtensor *sz, block_kind k, int which_pe, 
cannam@95: 		     INT *coords);
cannam@95: INT XM(total_block)(const dtensor *sz, block_kind k, int which_pe);
cannam@95: int XM(is_local_after)(int dim, const dtensor *sz, block_kind k);
cannam@95: int XM(is_local)(const dtensor *sz, block_kind k);
cannam@95: int XM(is_block1d)(const dtensor *sz, block_kind k);
cannam@95: 
cannam@95: /* choose-radix.c */
cannam@95: INT XM(choose_radix)(ddim d, int n_pes, unsigned flags, int sign,
cannam@95:                      INT rblock[2], INT mblock[2]);
cannam@95: 
cannam@95: /***********************************************************************/
cannam@95: /* any_true.c */
cannam@95: int XM(any_true)(int condition, MPI_Comm comm);
cannam@95: int XM(md5_equal)(md5 m, MPI_Comm comm);
cannam@95: 
cannam@95: /* conf.c */
cannam@95: void XM(conf_standard)(planner *p);
cannam@95: 
cannam@95: /***********************************************************************/
cannam@95: /* rearrange.c */
cannam@95: 
cannam@95: /* Different ways to rearrange the vector dimension vn during transposition,
cannam@95:    reflecting different tradeoffs between ease of transposition and
cannam@95:    contiguity during the subsequent DFTs.
cannam@95: 
cannam@95:    TODO: can we pare this down to CONTIG and DISCONTIG, at least
cannam@95:    in MEASURE mode?  SQUARE_MIDDLE is also used for 1d destroy-input DFTs. */
cannam@95: typedef enum {
cannam@95:      CONTIG = 0, /* vn x 1: make subsequent DFTs contiguous */
cannam@95:      DISCONTIG, /* P x (vn/P) for P processes */
cannam@95:      SQUARE_BEFORE, /* try to get square transpose at beginning */
cannam@95:      SQUARE_MIDDLE, /* try to get square transpose in the middle */
cannam@95:      SQUARE_AFTER /* try to get square transpose at end */
cannam@95: } rearrangement;
cannam@95: 
cannam@95: /* skipping SQUARE_AFTER since it doesn't seem to offer any advantage
cannam@95:    over SQUARE_BEFORE */
cannam@95: #define FORALL_REARRANGE(rearrange) for (rearrange = CONTIG; rearrange <= SQUARE_MIDDLE; rearrange = (rearrangement) (((int) rearrange) + 1))
cannam@95: 
cannam@95: int XM(rearrange_applicable)(rearrangement rearrange, 
cannam@95: 			     ddim dim0, INT vn, int n_pes);
cannam@95: INT XM(rearrange_ny)(rearrangement rearrange, ddim dim0, INT vn, int n_pes);
cannam@95: 
cannam@95: /***********************************************************************/
cannam@95: 
cannam@95: #endif /* __IFFTW_MPI_H__ */
cannam@95: