Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.5/mpi/dft-rank1-bigvec.c @ 43:5ea0608b923f

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Current zlib source
author Chris Cannam
date Tue, 18 Oct 2016 14:33:52 +0100
parents 2cd0e3b3e1fd
children
rev   line source
Chris@42 1 /*
Chris@42 2 * Copyright (c) 2003, 2007-14 Matteo Frigo
Chris@42 3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
Chris@42 4 *
Chris@42 5 * This program is free software; you can redistribute it and/or modify
Chris@42 6 * it under the terms of the GNU General Public License as published by
Chris@42 7 * the Free Software Foundation; either version 2 of the License, or
Chris@42 8 * (at your option) any later version.
Chris@42 9 *
Chris@42 10 * This program is distributed in the hope that it will be useful,
Chris@42 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@42 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
Chris@42 13 * GNU General Public License for more details.
Chris@42 14 *
Chris@42 15 * You should have received a copy of the GNU General Public License
Chris@42 16 * along with this program; if not, write to the Free Software
Chris@42 17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
Chris@42 18 *
Chris@42 19 */
Chris@42 20
Chris@42 21 /* Complex DFTs of rank == 1 when the vector length vn is >= # processes.
Chris@42 22 In this case, we don't need to use a six-step type algorithm, and can
Chris@42 23 instead transpose the DFT dimension with the vector dimension to
Chris@42 24 make the DFT local. */
Chris@42 25
Chris@42 26 #include "mpi-dft.h"
Chris@42 27 #include "mpi-transpose.h"
Chris@42 28 #include "dft.h"
Chris@42 29
Chris@42 30 typedef struct {
Chris@42 31 solver super;
Chris@42 32 int preserve_input; /* preserve input even if DESTROY_INPUT was passed */
Chris@42 33 rearrangement rearrange;
Chris@42 34 } S;
Chris@42 35
Chris@42 36 typedef struct {
Chris@42 37 plan_mpi_dft super;
Chris@42 38
Chris@42 39 plan *cldt_before, *cld, *cldt_after;
Chris@42 40 INT roff, ioff;
Chris@42 41 int preserve_input;
Chris@42 42 rearrangement rearrange;
Chris@42 43 } P;
Chris@42 44
Chris@42 45 static void apply(const plan *ego_, R *I, R *O)
Chris@42 46 {
Chris@42 47 const P *ego = (const P *) ego_;
Chris@42 48 plan_dft *cld;
Chris@42 49 plan_rdft *cldt_before, *cldt_after;
Chris@42 50 INT roff = ego->roff, ioff = ego->ioff;
Chris@42 51
Chris@42 52 /* global transpose */
Chris@42 53 cldt_before = (plan_rdft *) ego->cldt_before;
Chris@42 54 cldt_before->apply(ego->cldt_before, I, O);
Chris@42 55
Chris@42 56 if (ego->preserve_input) I = O;
Chris@42 57
Chris@42 58 /* 1d DFT(s) */
Chris@42 59 cld = (plan_dft *) ego->cld;
Chris@42 60 cld->apply(ego->cld, O+roff, O+ioff, I+roff, I+ioff);
Chris@42 61
Chris@42 62 /* global transpose */
Chris@42 63 cldt_after = (plan_rdft *) ego->cldt_after;
Chris@42 64 cldt_after->apply(ego->cldt_after, I, O);
Chris@42 65 }
Chris@42 66
Chris@42 67 static int applicable(const S *ego, const problem *p_,
Chris@42 68 const planner *plnr)
Chris@42 69 {
Chris@42 70 const problem_mpi_dft *p = (const problem_mpi_dft *) p_;
Chris@42 71 int n_pes;
Chris@42 72 MPI_Comm_size(p->comm, &n_pes);
Chris@42 73 return (1
Chris@42 74 && p->sz->rnk == 1
Chris@42 75 && !(p->flags & ~RANK1_BIGVEC_ONLY)
Chris@42 76 && (!ego->preserve_input || (!NO_DESTROY_INPUTP(plnr)
Chris@42 77 && p->I != p->O))
Chris@42 78 && (p->vn >= n_pes /* TODO: relax this, using more memory? */
Chris@42 79 || (p->flags & RANK1_BIGVEC_ONLY))
Chris@42 80
Chris@42 81 && XM(rearrange_applicable)(ego->rearrange,
Chris@42 82 p->sz->dims[0], p->vn, n_pes)
Chris@42 83
Chris@42 84 && (!NO_SLOWP(plnr) /* slow if dft-serial is applicable */
Chris@42 85 || !XM(dft_serial_applicable)(p))
Chris@42 86 );
Chris@42 87 }
Chris@42 88
Chris@42 89 static void awake(plan *ego_, enum wakefulness wakefulness)
Chris@42 90 {
Chris@42 91 P *ego = (P *) ego_;
Chris@42 92 X(plan_awake)(ego->cldt_before, wakefulness);
Chris@42 93 X(plan_awake)(ego->cld, wakefulness);
Chris@42 94 X(plan_awake)(ego->cldt_after, wakefulness);
Chris@42 95 }
Chris@42 96
Chris@42 97 static void destroy(plan *ego_)
Chris@42 98 {
Chris@42 99 P *ego = (P *) ego_;
Chris@42 100 X(plan_destroy_internal)(ego->cldt_after);
Chris@42 101 X(plan_destroy_internal)(ego->cld);
Chris@42 102 X(plan_destroy_internal)(ego->cldt_before);
Chris@42 103 }
Chris@42 104
Chris@42 105 static void print(const plan *ego_, printer *p)
Chris@42 106 {
Chris@42 107 const P *ego = (const P *) ego_;
Chris@42 108 const char descrip[][16] = { "contig", "discontig", "square-after",
Chris@42 109 "square-middle", "square-before" };
Chris@42 110 p->print(p, "(mpi-dft-rank1-bigvec/%s%s %(%p%) %(%p%) %(%p%))",
Chris@42 111 descrip[ego->rearrange], ego->preserve_input==2 ?"/p":"",
Chris@42 112 ego->cldt_before, ego->cld, ego->cldt_after);
Chris@42 113 }
Chris@42 114
Chris@42 115 static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
Chris@42 116 {
Chris@42 117 const S *ego = (const S *) ego_;
Chris@42 118 const problem_mpi_dft *p;
Chris@42 119 P *pln;
Chris@42 120 plan *cld = 0, *cldt_before = 0, *cldt_after = 0;
Chris@42 121 R *ri, *ii, *ro, *io, *I, *O;
Chris@42 122 INT yblock, yb, nx, ny, vn;
Chris@42 123 int my_pe, n_pes;
Chris@42 124 static const plan_adt padt = {
Chris@42 125 XM(dft_solve), awake, print, destroy
Chris@42 126 };
Chris@42 127
Chris@42 128 UNUSED(ego);
Chris@42 129
Chris@42 130 if (!applicable(ego, p_, plnr))
Chris@42 131 return (plan *) 0;
Chris@42 132
Chris@42 133 p = (const problem_mpi_dft *) p_;
Chris@42 134
Chris@42 135 MPI_Comm_rank(p->comm, &my_pe);
Chris@42 136 MPI_Comm_size(p->comm, &n_pes);
Chris@42 137
Chris@42 138 nx = p->sz->dims[0].n;
Chris@42 139 if (!(ny = XM(rearrange_ny)(ego->rearrange, p->sz->dims[0],p->vn,n_pes)))
Chris@42 140 return (plan *) 0;
Chris@42 141 vn = p->vn / ny;
Chris@42 142 A(ny * vn == p->vn);
Chris@42 143
Chris@42 144 yblock = XM(default_block)(ny, n_pes);
Chris@42 145 cldt_before = X(mkplan_d)(plnr,
Chris@42 146 XM(mkproblem_transpose)(
Chris@42 147 nx, ny, vn*2,
Chris@42 148 I = p->I, O = p->O,
Chris@42 149 p->sz->dims[0].b[IB], yblock,
Chris@42 150 p->comm, 0));
Chris@42 151 if (XM(any_true)(!cldt_before, p->comm)) goto nada;
Chris@42 152 if (ego->preserve_input || NO_DESTROY_INPUTP(plnr)) { I = O; }
Chris@42 153
Chris@42 154 X(extract_reim)(p->sign, I, &ri, &ii);
Chris@42 155 X(extract_reim)(p->sign, O, &ro, &io);
Chris@42 156
Chris@42 157 yb = XM(block)(ny, yblock, my_pe);
Chris@42 158 cld = X(mkplan_d)(plnr,
Chris@42 159 X(mkproblem_dft_d)(X(mktensor_1d)(nx, vn*2, vn*2),
Chris@42 160 X(mktensor_2d)(yb, vn*2*nx, vn*2*nx,
Chris@42 161 vn, 2, 2),
Chris@42 162 ro, io, ri, ii));
Chris@42 163 if (XM(any_true)(!cld, p->comm)) goto nada;
Chris@42 164
Chris@42 165 cldt_after = X(mkplan_d)(plnr,
Chris@42 166 XM(mkproblem_transpose)(
Chris@42 167 ny, nx, vn*2,
Chris@42 168 I, O,
Chris@42 169 yblock, p->sz->dims[0].b[OB],
Chris@42 170 p->comm, 0));
Chris@42 171 if (XM(any_true)(!cldt_after, p->comm)) goto nada;
Chris@42 172
Chris@42 173 pln = MKPLAN_MPI_DFT(P, &padt, apply);
Chris@42 174
Chris@42 175 pln->cldt_before = cldt_before;
Chris@42 176 pln->cld = cld;
Chris@42 177 pln->cldt_after = cldt_after;
Chris@42 178 pln->preserve_input = ego->preserve_input ? 2 : NO_DESTROY_INPUTP(plnr);
Chris@42 179 pln->roff = ro - p->O;
Chris@42 180 pln->ioff = io - p->O;
Chris@42 181 pln->rearrange = ego->rearrange;
Chris@42 182
Chris@42 183 X(ops_add)(&cldt_before->ops, &cld->ops, &pln->super.super.ops);
Chris@42 184 X(ops_add2)(&cldt_after->ops, &pln->super.super.ops);
Chris@42 185
Chris@42 186 return &(pln->super.super);
Chris@42 187
Chris@42 188 nada:
Chris@42 189 X(plan_destroy_internal)(cldt_after);
Chris@42 190 X(plan_destroy_internal)(cld);
Chris@42 191 X(plan_destroy_internal)(cldt_before);
Chris@42 192 return (plan *) 0;
Chris@42 193 }
Chris@42 194
Chris@42 195 static solver *mksolver(rearrangement rearrange, int preserve_input)
Chris@42 196 {
Chris@42 197 static const solver_adt sadt = { PROBLEM_MPI_DFT, mkplan, 0 };
Chris@42 198 S *slv = MKSOLVER(S, &sadt);
Chris@42 199 slv->rearrange = rearrange;
Chris@42 200 slv->preserve_input = preserve_input;
Chris@42 201 return &(slv->super);
Chris@42 202 }
Chris@42 203
Chris@42 204 void XM(dft_rank1_bigvec_register)(planner *p)
Chris@42 205 {
Chris@42 206 rearrangement rearrange;
Chris@42 207 int preserve_input;
Chris@42 208 FORALL_REARRANGE(rearrange)
Chris@42 209 for (preserve_input = 0; preserve_input <= 1; ++preserve_input)
Chris@42 210 REGISTER_SOLVER(p, mksolver(rearrange, preserve_input));
Chris@42 211 }