Chris@10: /*
Chris@10:  * Copyright (c) 2003, 2007-11 Matteo Frigo
Chris@10:  * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
Chris@10:  *
Chris@10:  * This program is free software; you can redistribute it and/or modify
Chris@10:  * it under the terms of the GNU General Public License as published by
Chris@10:  * the Free Software Foundation; either version 2 of the License, or
Chris@10:  * (at your option) any later version.
Chris@10:  *
Chris@10:  * This program is distributed in the hope that it will be useful,
Chris@10:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@10:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
Chris@10:  * GNU General Public License for more details.
Chris@10:  *
Chris@10:  * You should have received a copy of the GNU General Public License
Chris@10:  * along with this program; if not, write to the Free Software
Chris@10:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
Chris@10:  *
Chris@10:  */
Chris@10: 
Chris@10: /* FFTW-MPI internal header file */
Chris@10: #ifndef __IFFTW_MPI_H__
Chris@10: #define __IFFTW_MPI_H__
Chris@10: 
Chris@10: #include "ifftw.h"
Chris@10: #include "rdft.h"
Chris@10: 
Chris@10: #include <mpi.h>
Chris@10: 
Chris@10: /* mpi problem flags: problem-dependent meaning, but in general
Chris@10:    SCRAMBLED means some reordering *within* the dimensions, while
Chris@10:    TRANSPOSED means some reordering *of* the dimensions */
Chris@10: #define SCRAMBLED_IN (1 << 0)
Chris@10: #define SCRAMBLED_OUT (1 << 1)
Chris@10: #define TRANSPOSED_IN (1 << 2)
Chris@10: #define TRANSPOSED_OUT (1 << 3)
Chris@10: #define RANK1_BIGVEC_ONLY (1 << 4) /* for rank=1, allow only bigvec solver */
Chris@10: 
Chris@10: #define ONLY_SCRAMBLEDP(flags) (!((flags) & ~(SCRAMBLED_IN|SCRAMBLED_OUT)))
Chris@10: #define ONLY_TRANSPOSEDP(flags) (!((flags) & ~(TRANSPOSED_IN|TRANSPOSED_OUT)))
Chris@10: 
Chris@10: #if defined(FFTW_SINGLE)
Chris@10: #  define FFTW_MPI_TYPE MPI_FLOAT
Chris@10: #elif defined(FFTW_LDOUBLE)
Chris@10: #  define FFTW_MPI_TYPE MPI_LONG_DOUBLE
Chris@10: #elif defined(FFTW_QUAD)
Chris@10: #  error MPI quad-precision type is unknown
Chris@10: #else
Chris@10: #  define FFTW_MPI_TYPE MPI_DOUBLE
Chris@10: #endif
Chris@10: 
Chris@10: /* all fftw-mpi identifiers start with fftw_mpi (or fftwf_mpi etc.) */
Chris@10: #define XM(name) X(CONCAT(mpi_, name))
Chris@10: 
Chris@10: /***********************************************************************/
Chris@10: /* block distributions */
Chris@10: 
Chris@10: /* a distributed dimension of length n with input and output block
Chris@10:    sizes ib and ob, respectively. */
Chris@10: typedef enum { IB = 0, OB } block_kind;
Chris@10: typedef struct {
Chris@10:      INT n;
Chris@10:      INT b[2]; /* b[IB], b[OB] */
Chris@10: } ddim;
Chris@10: 
Chris@10: /* Loop over k in {IB, OB}.  Note: need explicit casts for C++. */
Chris@10: #define FORALL_BLOCK_KIND(k) for (k = IB; k <= OB; k = (block_kind) (((int) k) + 1))
Chris@10: 
Chris@10: /* unlike tensors in the serial FFTW, the ordering of the dtensor
Chris@10:    dimensions matters - both the array and the block layout are
Chris@10:    row-major order. */
Chris@10: typedef struct {
Chris@10:      int rnk;
Chris@10: #if defined(STRUCT_HACK_KR)
Chris@10:      ddim dims[1];
Chris@10: #elif defined(STRUCT_HACK_C99)
Chris@10:      ddim dims[];
Chris@10: #else
Chris@10:      ddim *dims;
Chris@10: #endif
Chris@10: } dtensor;
Chris@10: 
Chris@10: 
Chris@10: /* dtensor.c: */
Chris@10: dtensor *XM(mkdtensor)(int rnk);
Chris@10: void XM(dtensor_destroy)(dtensor *sz);
Chris@10: dtensor *XM(dtensor_copy)(const dtensor *sz);
Chris@10: dtensor *XM(dtensor_canonical)(const dtensor *sz, int compress);
Chris@10: int XM(dtensor_validp)(const dtensor *sz);
Chris@10: void XM(dtensor_md5)(md5 *p, const dtensor *t);
Chris@10: void XM(dtensor_print)(const dtensor *t, printer *p);
Chris@10: 
Chris@10: /* block.c: */
Chris@10: 
Chris@10: /* for a single distributed dimension: */
Chris@10: INT XM(num_blocks)(INT n, INT block);
Chris@10: int XM(num_blocks_ok)(INT n, INT block, MPI_Comm comm);
Chris@10: INT XM(default_block)(INT n, int n_pes);
Chris@10: INT XM(block)(INT n, INT block, int which_block);
Chris@10: 
Chris@10: /* for multiple distributed dimensions: */
Chris@10: INT XM(num_blocks_total)(const dtensor *sz, block_kind k);
Chris@10: int XM(idle_process)(const dtensor *sz, block_kind k, int which_pe);
Chris@10: void XM(block_coords)(const dtensor *sz, block_kind k, int which_pe, 
Chris@10: 		     INT *coords);
Chris@10: INT XM(total_block)(const dtensor *sz, block_kind k, int which_pe);
Chris@10: int XM(is_local_after)(int dim, const dtensor *sz, block_kind k);
Chris@10: int XM(is_local)(const dtensor *sz, block_kind k);
Chris@10: int XM(is_block1d)(const dtensor *sz, block_kind k);
Chris@10: 
Chris@10: /* choose-radix.c */
Chris@10: INT XM(choose_radix)(ddim d, int n_pes, unsigned flags, int sign,
Chris@10:                      INT rblock[2], INT mblock[2]);
Chris@10: 
Chris@10: /***********************************************************************/
Chris@10: /* any_true.c */
Chris@10: int XM(any_true)(int condition, MPI_Comm comm);
Chris@10: int XM(md5_equal)(md5 m, MPI_Comm comm);
Chris@10: 
Chris@10: /* conf.c */
Chris@10: void XM(conf_standard)(planner *p);
Chris@10: 
Chris@10: /***********************************************************************/
Chris@10: /* rearrange.c */
Chris@10: 
Chris@10: /* Different ways to rearrange the vector dimension vn during transposition,
Chris@10:    reflecting different tradeoffs between ease of transposition and
Chris@10:    contiguity during the subsequent DFTs.
Chris@10: 
Chris@10:    TODO: can we pare this down to CONTIG and DISCONTIG, at least
Chris@10:    in MEASURE mode?  SQUARE_MIDDLE is also used for 1d destroy-input DFTs. */
Chris@10: typedef enum {
Chris@10:      CONTIG = 0, /* vn x 1: make subsequent DFTs contiguous */
Chris@10:      DISCONTIG, /* P x (vn/P) for P processes */
Chris@10:      SQUARE_BEFORE, /* try to get square transpose at beginning */
Chris@10:      SQUARE_MIDDLE, /* try to get square transpose in the middle */
Chris@10:      SQUARE_AFTER /* try to get square transpose at end */
Chris@10: } rearrangement;
Chris@10: 
Chris@10: /* skipping SQUARE_AFTER since it doesn't seem to offer any advantage
Chris@10:    over SQUARE_BEFORE */
Chris@10: #define FORALL_REARRANGE(rearrange) for (rearrange = CONTIG; rearrange <= SQUARE_MIDDLE; rearrange = (rearrangement) (((int) rearrange) + 1))
Chris@10: 
Chris@10: int XM(rearrange_applicable)(rearrangement rearrange, 
Chris@10: 			     ddim dim0, INT vn, int n_pes);
Chris@10: INT XM(rearrange_ny)(rearrangement rearrange, ddim dim0, INT vn, int n_pes);
Chris@10: 
Chris@10: /***********************************************************************/
Chris@10: 
Chris@10: #endif /* __IFFTW_MPI_H__ */
Chris@10: