Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.5/threads/ct.c @ 42:2cd0e3b3e1fd

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Current fftw source
author Chris Cannam
date Tue, 18 Oct 2016 13:40:26 +0100
parents
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
Chris@42 22 #include "threads.h"
Chris@42 23
Chris@42 24 typedef struct {
Chris@42 25 plan_dft super;
Chris@42 26 plan *cld;
Chris@42 27 plan **cldws;
Chris@42 28 int nthr;
Chris@42 29 INT r;
Chris@42 30 } P;
Chris@42 31
Chris@42 32 typedef struct {
Chris@42 33 plan **cldws;
Chris@42 34 R *r, *i;
Chris@42 35 } PD;
Chris@42 36
Chris@42 37 static void *spawn_apply(spawn_data *d)
Chris@42 38 {
Chris@42 39 PD *ego = (PD *) d->data;
Chris@42 40 INT thr_num = d->thr_num;
Chris@42 41
Chris@42 42 plan_dftw *cldw = (plan_dftw *) (ego->cldws[thr_num]);
Chris@42 43 cldw->apply((plan *) cldw, ego->r, ego->i);
Chris@42 44 return 0;
Chris@42 45 }
Chris@42 46
Chris@42 47 static void apply_dit(const plan *ego_, R *ri, R *ii, R *ro, R *io)
Chris@42 48 {
Chris@42 49 const P *ego = (const P *) ego_;
Chris@42 50 plan_dft *cld;
Chris@42 51
Chris@42 52 cld = (plan_dft *) ego->cld;
Chris@42 53 cld->apply(ego->cld, ri, ii, ro, io);
Chris@42 54
Chris@42 55 {
Chris@42 56 PD d;
Chris@42 57
Chris@42 58 d.r = ro; d.i = io;
Chris@42 59 d.cldws = ego->cldws;
Chris@42 60
Chris@42 61 X(spawn_loop)(ego->nthr, ego->nthr, spawn_apply, (void*)&d);
Chris@42 62 }
Chris@42 63 }
Chris@42 64
Chris@42 65 static void apply_dif(const plan *ego_, R *ri, R *ii, R *ro, R *io)
Chris@42 66 {
Chris@42 67 const P *ego = (const P *) ego_;
Chris@42 68 plan_dft *cld;
Chris@42 69
Chris@42 70 {
Chris@42 71 PD d;
Chris@42 72
Chris@42 73 d.r = ri; d.i = ii;
Chris@42 74 d.cldws = ego->cldws;
Chris@42 75
Chris@42 76 X(spawn_loop)(ego->nthr, ego->nthr, spawn_apply, (void*)&d);
Chris@42 77 }
Chris@42 78
Chris@42 79 cld = (plan_dft *) ego->cld;
Chris@42 80 cld->apply(ego->cld, ri, ii, ro, io);
Chris@42 81 }
Chris@42 82
Chris@42 83 static void awake(plan *ego_, enum wakefulness wakefulness)
Chris@42 84 {
Chris@42 85 P *ego = (P *) ego_;
Chris@42 86 int i;
Chris@42 87 X(plan_awake)(ego->cld, wakefulness);
Chris@42 88 for (i = 0; i < ego->nthr; ++i)
Chris@42 89 X(plan_awake)(ego->cldws[i], wakefulness);
Chris@42 90 }
Chris@42 91
Chris@42 92 static void destroy(plan *ego_)
Chris@42 93 {
Chris@42 94 P *ego = (P *) ego_;
Chris@42 95 int i;
Chris@42 96 X(plan_destroy_internal)(ego->cld);
Chris@42 97 for (i = 0; i < ego->nthr; ++i)
Chris@42 98 X(plan_destroy_internal)(ego->cldws[i]);
Chris@42 99 X(ifree)(ego->cldws);
Chris@42 100 }
Chris@42 101
Chris@42 102 static void print(const plan *ego_, printer *p)
Chris@42 103 {
Chris@42 104 const P *ego = (const P *) ego_;
Chris@42 105 int i;
Chris@42 106 p->print(p, "(dft-thr-ct-%s-x%d/%D",
Chris@42 107 ego->super.apply == apply_dit ? "dit" : "dif",
Chris@42 108 ego->nthr, ego->r);
Chris@42 109 for (i = 0; i < ego->nthr; ++i)
Chris@42 110 if (i == 0 || (ego->cldws[i] != ego->cldws[i-1] &&
Chris@42 111 (i <= 1 || ego->cldws[i] != ego->cldws[i-2])))
Chris@42 112 p->print(p, "%(%p%)", ego->cldws[i]);
Chris@42 113 p->print(p, "%(%p%))", ego->cld);
Chris@42 114 }
Chris@42 115
Chris@42 116 static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
Chris@42 117 {
Chris@42 118 const ct_solver *ego = (const ct_solver *) ego_;
Chris@42 119 const problem_dft *p;
Chris@42 120 P *pln = 0;
Chris@42 121 plan *cld = 0, **cldws = 0;
Chris@42 122 INT n, r, m, v, ivs, ovs;
Chris@42 123 INT block_size;
Chris@42 124 int i, nthr, plnr_nthr_save;
Chris@42 125 iodim *d;
Chris@42 126
Chris@42 127 static const plan_adt padt = {
Chris@42 128 X(dft_solve), awake, print, destroy
Chris@42 129 };
Chris@42 130
Chris@42 131 if (plnr->nthr <= 1 || !X(ct_applicable)(ego, p_, plnr))
Chris@42 132 return (plan *) 0;
Chris@42 133
Chris@42 134 p = (const problem_dft *) p_;
Chris@42 135 d = p->sz->dims;
Chris@42 136 n = d[0].n;
Chris@42 137 r = X(choose_radix)(ego->r, n);
Chris@42 138 m = n / r;
Chris@42 139
Chris@42 140 X(tensor_tornk1)(p->vecsz, &v, &ivs, &ovs);
Chris@42 141
Chris@42 142 block_size = (m + plnr->nthr - 1) / plnr->nthr;
Chris@42 143 nthr = (int)((m + block_size - 1) / block_size);
Chris@42 144 plnr_nthr_save = plnr->nthr;
Chris@42 145 plnr->nthr = (plnr->nthr + nthr - 1) / nthr;
Chris@42 146
Chris@42 147 cldws = (plan **) MALLOC(sizeof(plan *) * nthr, PLANS);
Chris@42 148 for (i = 0; i < nthr; ++i) cldws[i] = (plan *) 0;
Chris@42 149
Chris@42 150 switch (ego->dec) {
Chris@42 151 case DECDIT:
Chris@42 152 {
Chris@42 153 for (i = 0; i < nthr; ++i) {
Chris@42 154 cldws[i] = ego->mkcldw(ego,
Chris@42 155 r, m * d[0].os, m * d[0].os,
Chris@42 156 m, d[0].os,
Chris@42 157 v, ovs, ovs,
Chris@42 158 i*block_size,
Chris@42 159 (i == nthr - 1) ?
Chris@42 160 (m - i*block_size) : block_size,
Chris@42 161 p->ro, p->io, plnr);
Chris@42 162 if (!cldws[i]) goto nada;
Chris@42 163 }
Chris@42 164
Chris@42 165 plnr->nthr = plnr_nthr_save;
Chris@42 166
Chris@42 167 cld = X(mkplan_d)(plnr,
Chris@42 168 X(mkproblem_dft_d)(
Chris@42 169 X(mktensor_1d)(m, r * d[0].is, d[0].os),
Chris@42 170 X(mktensor_2d)(r, d[0].is, m * d[0].os,
Chris@42 171 v, ivs, ovs),
Chris@42 172 p->ri, p->ii, p->ro, p->io)
Chris@42 173 );
Chris@42 174 if (!cld) goto nada;
Chris@42 175
Chris@42 176 pln = MKPLAN_DFT(P, &padt, apply_dit);
Chris@42 177 break;
Chris@42 178 }
Chris@42 179 case DECDIF:
Chris@42 180 case DECDIF+TRANSPOSE:
Chris@42 181 {
Chris@42 182 INT cors, covs; /* cldw ors, ovs */
Chris@42 183 if (ego->dec == DECDIF+TRANSPOSE) {
Chris@42 184 cors = ivs;
Chris@42 185 covs = m * d[0].is;
Chris@42 186 /* ensure that we generate well-formed dftw subproblems */
Chris@42 187 /* FIXME: too conservative */
Chris@42 188 if (!(1
Chris@42 189 && r == v
Chris@42 190 && d[0].is == r * cors))
Chris@42 191 goto nada;
Chris@42 192
Chris@42 193 /* FIXME: allow in-place only for now, like in
Chris@42 194 fftw-3.[01] */
Chris@42 195 if (!(1
Chris@42 196 && p->ri == p->ro
Chris@42 197 && d[0].is == r * d[0].os
Chris@42 198 && cors == d[0].os
Chris@42 199 && covs == ovs
Chris@42 200 ))
Chris@42 201 goto nada;
Chris@42 202 } else {
Chris@42 203 cors = m * d[0].is;
Chris@42 204 covs = ivs;
Chris@42 205 }
Chris@42 206
Chris@42 207 for (i = 0; i < nthr; ++i) {
Chris@42 208 cldws[i] = ego->mkcldw(ego,
Chris@42 209 r, m * d[0].is, cors,
Chris@42 210 m, d[0].is,
Chris@42 211 v, ivs, covs,
Chris@42 212 i*block_size,
Chris@42 213 (i == nthr - 1) ?
Chris@42 214 (m - i*block_size) : block_size,
Chris@42 215 p->ri, p->ii, plnr);
Chris@42 216 if (!cldws[i]) goto nada;
Chris@42 217 }
Chris@42 218
Chris@42 219 plnr->nthr = plnr_nthr_save;
Chris@42 220
Chris@42 221 cld = X(mkplan_d)(plnr,
Chris@42 222 X(mkproblem_dft_d)(
Chris@42 223 X(mktensor_1d)(m, d[0].is, r * d[0].os),
Chris@42 224 X(mktensor_2d)(r, cors, d[0].os,
Chris@42 225 v, covs, ovs),
Chris@42 226 p->ri, p->ii, p->ro, p->io)
Chris@42 227 );
Chris@42 228 if (!cld) goto nada;
Chris@42 229
Chris@42 230 pln = MKPLAN_DFT(P, &padt, apply_dif);
Chris@42 231 break;
Chris@42 232 }
Chris@42 233
Chris@42 234 default: A(0);
Chris@42 235
Chris@42 236 }
Chris@42 237
Chris@42 238 pln->cld = cld;
Chris@42 239 pln->cldws = cldws;
Chris@42 240 pln->nthr = nthr;
Chris@42 241 pln->r = r;
Chris@42 242 X(ops_zero)(&pln->super.super.ops);
Chris@42 243 for (i = 0; i < nthr; ++i) {
Chris@42 244 X(ops_add2)(&cldws[i]->ops, &pln->super.super.ops);
Chris@42 245 pln->super.super.could_prune_now_p |= cldws[i]->could_prune_now_p;
Chris@42 246 }
Chris@42 247 X(ops_add2)(&cld->ops, &pln->super.super.ops);
Chris@42 248 return &(pln->super.super);
Chris@42 249
Chris@42 250 nada:
Chris@42 251 if (cldws) {
Chris@42 252 for (i = 0; i < nthr; ++i)
Chris@42 253 X(plan_destroy_internal)(cldws[i]);
Chris@42 254 X(ifree)(cldws);
Chris@42 255 }
Chris@42 256 X(plan_destroy_internal)(cld);
Chris@42 257 return (plan *) 0;
Chris@42 258 }
Chris@42 259
Chris@42 260 ct_solver *X(mksolver_ct_threads)(size_t size, INT r, int dec,
Chris@42 261 ct_mkinferior mkcldw,
Chris@42 262 ct_force_vrecursion force_vrecursionp)
Chris@42 263 {
Chris@42 264 static const solver_adt sadt = { PROBLEM_DFT, mkplan, 0 };
Chris@42 265 ct_solver *slv = (ct_solver *) X(mksolver)(size, &sadt);
Chris@42 266 slv->r = r;
Chris@42 267 slv->dec = dec;
Chris@42 268 slv->mkcldw = mkcldw;
Chris@42 269 slv->force_vrecursionp = force_vrecursionp;
Chris@42 270 return slv;
Chris@42 271 }