Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.8/mpi/ifftw-mpi.h @ 82:d0c2a83c1364

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Add FFTW 3.3.8 source, and a Linux build
author Chris Cannam
date Tue, 19 Nov 2019 14:52:55 +0000
parents
children
rev   line source
Chris@82 1 /*
Chris@82 2 * Copyright (c) 2003, 2007-14 Matteo Frigo
Chris@82 3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
Chris@82 4 *
Chris@82 5 * This program is free software; you can redistribute it and/or modify
Chris@82 6 * it under the terms of the GNU General Public License as published by
Chris@82 7 * the Free Software Foundation; either version 2 of the License, or
Chris@82 8 * (at your option) any later version.
Chris@82 9 *
Chris@82 10 * This program is distributed in the hope that it will be useful,
Chris@82 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@82 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
Chris@82 13 * GNU General Public License for more details.
Chris@82 14 *
Chris@82 15 * You should have received a copy of the GNU General Public License
Chris@82 16 * along with this program; if not, write to the Free Software
Chris@82 17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
Chris@82 18 *
Chris@82 19 */
Chris@82 20
Chris@82 21 /* FFTW-MPI internal header file */
Chris@82 22 #ifndef __IFFTW_MPI_H__
Chris@82 23 #define __IFFTW_MPI_H__
Chris@82 24
Chris@82 25 #include "kernel/ifftw.h"
Chris@82 26 #include "rdft/rdft.h"
Chris@82 27
Chris@82 28 #include <mpi.h>
Chris@82 29
Chris@82 30 /* mpi problem flags: problem-dependent meaning, but in general
Chris@82 31 SCRAMBLED means some reordering *within* the dimensions, while
Chris@82 32 TRANSPOSED means some reordering *of* the dimensions */
Chris@82 33 #define SCRAMBLED_IN (1 << 0)
Chris@82 34 #define SCRAMBLED_OUT (1 << 1)
Chris@82 35 #define TRANSPOSED_IN (1 << 2)
Chris@82 36 #define TRANSPOSED_OUT (1 << 3)
Chris@82 37 #define RANK1_BIGVEC_ONLY (1 << 4) /* for rank=1, allow only bigvec solver */
Chris@82 38
Chris@82 39 #define ONLY_SCRAMBLEDP(flags) (!((flags) & ~(SCRAMBLED_IN|SCRAMBLED_OUT)))
Chris@82 40 #define ONLY_TRANSPOSEDP(flags) (!((flags) & ~(TRANSPOSED_IN|TRANSPOSED_OUT)))
Chris@82 41
Chris@82 42 #if defined(FFTW_SINGLE)
Chris@82 43 # define FFTW_MPI_TYPE MPI_FLOAT
Chris@82 44 #elif defined(FFTW_LDOUBLE)
Chris@82 45 # define FFTW_MPI_TYPE MPI_LONG_DOUBLE
Chris@82 46 #elif defined(FFTW_QUAD)
Chris@82 47 # error MPI quad-precision type is unknown
Chris@82 48 #else
Chris@82 49 # define FFTW_MPI_TYPE MPI_DOUBLE
Chris@82 50 #endif
Chris@82 51
Chris@82 52 /* all fftw-mpi identifiers start with fftw_mpi (or fftwf_mpi etc.) */
Chris@82 53 #define XM(name) X(CONCAT(mpi_, name))
Chris@82 54
Chris@82 55 /***********************************************************************/
Chris@82 56 /* block distributions */
Chris@82 57
Chris@82 58 /* a distributed dimension of length n with input and output block
Chris@82 59 sizes ib and ob, respectively. */
Chris@82 60 typedef enum { IB = 0, OB } block_kind;
Chris@82 61 typedef struct {
Chris@82 62 INT n;
Chris@82 63 INT b[2]; /* b[IB], b[OB] */
Chris@82 64 } ddim;
Chris@82 65
Chris@82 66 /* Loop over k in {IB, OB}. Note: need explicit casts for C++. */
Chris@82 67 #define FORALL_BLOCK_KIND(k) for (k = IB; k <= OB; k = (block_kind) (((int) k) + 1))
Chris@82 68
Chris@82 69 /* unlike tensors in the serial FFTW, the ordering of the dtensor
Chris@82 70 dimensions matters - both the array and the block layout are
Chris@82 71 row-major order. */
Chris@82 72 typedef struct {
Chris@82 73 int rnk;
Chris@82 74 #if defined(STRUCT_HACK_KR)
Chris@82 75 ddim dims[1];
Chris@82 76 #elif defined(STRUCT_HACK_C99)
Chris@82 77 ddim dims[];
Chris@82 78 #else
Chris@82 79 ddim *dims;
Chris@82 80 #endif
Chris@82 81 } dtensor;
Chris@82 82
Chris@82 83
Chris@82 84 /* dtensor.c: */
Chris@82 85 dtensor *XM(mkdtensor)(int rnk);
Chris@82 86 void XM(dtensor_destroy)(dtensor *sz);
Chris@82 87 dtensor *XM(dtensor_copy)(const dtensor *sz);
Chris@82 88 dtensor *XM(dtensor_canonical)(const dtensor *sz, int compress);
Chris@82 89 int XM(dtensor_validp)(const dtensor *sz);
Chris@82 90 void XM(dtensor_md5)(md5 *p, const dtensor *t);
Chris@82 91 void XM(dtensor_print)(const dtensor *t, printer *p);
Chris@82 92
Chris@82 93 /* block.c: */
Chris@82 94
Chris@82 95 /* for a single distributed dimension: */
Chris@82 96 INT XM(num_blocks)(INT n, INT block);
Chris@82 97 int XM(num_blocks_ok)(INT n, INT block, MPI_Comm comm);
Chris@82 98 INT XM(default_block)(INT n, int n_pes);
Chris@82 99 INT XM(block)(INT n, INT block, int which_block);
Chris@82 100
Chris@82 101 /* for multiple distributed dimensions: */
Chris@82 102 INT XM(num_blocks_total)(const dtensor *sz, block_kind k);
Chris@82 103 int XM(idle_process)(const dtensor *sz, block_kind k, int which_pe);
Chris@82 104 void XM(block_coords)(const dtensor *sz, block_kind k, int which_pe,
Chris@82 105 INT *coords);
Chris@82 106 INT XM(total_block)(const dtensor *sz, block_kind k, int which_pe);
Chris@82 107 int XM(is_local_after)(int dim, const dtensor *sz, block_kind k);
Chris@82 108 int XM(is_local)(const dtensor *sz, block_kind k);
Chris@82 109 int XM(is_block1d)(const dtensor *sz, block_kind k);
Chris@82 110
Chris@82 111 /* choose-radix.c */
Chris@82 112 INT XM(choose_radix)(ddim d, int n_pes, unsigned flags, int sign,
Chris@82 113 INT rblock[2], INT mblock[2]);
Chris@82 114
Chris@82 115 /***********************************************************************/
Chris@82 116 /* any_true.c */
Chris@82 117 int XM(any_true)(int condition, MPI_Comm comm);
Chris@82 118 int XM(md5_equal)(md5 m, MPI_Comm comm);
Chris@82 119
Chris@82 120 /* conf.c */
Chris@82 121 void XM(conf_standard)(planner *p);
Chris@82 122
Chris@82 123 /***********************************************************************/
Chris@82 124 /* rearrange.c */
Chris@82 125
Chris@82 126 /* Different ways to rearrange the vector dimension vn during transposition,
Chris@82 127 reflecting different tradeoffs between ease of transposition and
Chris@82 128 contiguity during the subsequent DFTs.
Chris@82 129
Chris@82 130 TODO: can we pare this down to CONTIG and DISCONTIG, at least
Chris@82 131 in MEASURE mode? SQUARE_MIDDLE is also used for 1d destroy-input DFTs. */
Chris@82 132 typedef enum {
Chris@82 133 CONTIG = 0, /* vn x 1: make subsequent DFTs contiguous */
Chris@82 134 DISCONTIG, /* P x (vn/P) for P processes */
Chris@82 135 SQUARE_BEFORE, /* try to get square transpose at beginning */
Chris@82 136 SQUARE_MIDDLE, /* try to get square transpose in the middle */
Chris@82 137 SQUARE_AFTER /* try to get square transpose at end */
Chris@82 138 } rearrangement;
Chris@82 139
Chris@82 140 /* skipping SQUARE_AFTER since it doesn't seem to offer any advantage
Chris@82 141 over SQUARE_BEFORE */
Chris@82 142 #define FORALL_REARRANGE(rearrange) for (rearrange = CONTIG; rearrange <= SQUARE_MIDDLE; rearrange = (rearrangement) (((int) rearrange) + 1))
Chris@82 143
Chris@82 144 int XM(rearrange_applicable)(rearrangement rearrange,
Chris@82 145 ddim dim0, INT vn, int n_pes);
Chris@82 146 INT XM(rearrange_ny)(rearrangement rearrange, ddim dim0, INT vn, int n_pes);
Chris@82 147
Chris@82 148 /***********************************************************************/
Chris@82 149
Chris@82 150 #endif /* __IFFTW_MPI_H__ */
Chris@82 151