cannam@127: /* cannam@127: * Copyright (c) 2003, 2007-14 Matteo Frigo cannam@127: * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology cannam@127: * cannam@127: * This program is free software; you can redistribute it and/or modify cannam@127: * it under the terms of the GNU General Public License as published by cannam@127: * the Free Software Foundation; either version 2 of the License, or cannam@127: * (at your option) any later version. cannam@127: * cannam@127: * This program is distributed in the hope that it will be useful, cannam@127: * but WITHOUT ANY WARRANTY; without even the implied warranty of cannam@127: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the cannam@127: * GNU General Public License for more details. cannam@127: * cannam@127: * You should have received a copy of the GNU General Public License cannam@127: * along with this program; if not, write to the Free Software cannam@127: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA cannam@127: * cannam@127: */ cannam@127: cannam@127: #include "mpi-rdft.h" cannam@127: cannam@127: static void destroy(problem *ego_) cannam@127: { cannam@127: problem_mpi_rdft *ego = (problem_mpi_rdft *) ego_; cannam@127: XM(dtensor_destroy)(ego->sz); cannam@127: MPI_Comm_free(&ego->comm); cannam@127: #if !defined(STRUCT_HACK_C99) && !defined(STRUCT_HACK_KR) cannam@127: X(ifree0)(ego->kind); cannam@127: #endif cannam@127: X(ifree)(ego_); cannam@127: } cannam@127: cannam@127: static void hash(const problem *p_, md5 *m) cannam@127: { cannam@127: const problem_mpi_rdft *p = (const problem_mpi_rdft *) p_; cannam@127: int i; cannam@127: X(md5puts)(m, "mpi-dft"); cannam@127: X(md5int)(m, p->I == p->O); cannam@127: /* don't include alignment -- may differ between processes cannam@127: X(md5int)(m, X(ialignment_of)(p->I)); cannam@127: X(md5int)(m, X(ialignment_of)(p->O)); cannam@127: ... note that applicability of MPI plans does not depend cannam@127: on alignment (although optimality may, in principle). */ cannam@127: XM(dtensor_md5)(m, p->sz); cannam@127: X(md5INT)(m, p->vn); cannam@127: for (i = 0; i < p->sz->rnk; ++i) cannam@127: X(md5int)(m, p->kind[i]); cannam@127: X(md5int)(m, p->flags); cannam@127: MPI_Comm_size(p->comm, &i); X(md5int)(m, i); cannam@127: A(XM(md5_equal)(*m, p->comm)); cannam@127: } cannam@127: cannam@127: static void print(const problem *ego_, printer *p) cannam@127: { cannam@127: const problem_mpi_rdft *ego = (const problem_mpi_rdft *) ego_; cannam@127: int i; cannam@127: p->print(p, "(mpi-rdft %d %d %d ", cannam@127: ego->I == ego->O, cannam@127: X(ialignment_of)(ego->I), cannam@127: X(ialignment_of)(ego->O)); cannam@127: XM(dtensor_print)(ego->sz, p); cannam@127: for (i = 0; i < ego->sz->rnk; ++i) cannam@127: p->print(p, " %d", (int)ego->kind[i]); cannam@127: p->print(p, " %D %d", ego->vn, ego->flags); cannam@127: MPI_Comm_size(ego->comm, &i); p->print(p, " %d)", i); cannam@127: } cannam@127: cannam@127: static void zero(const problem *ego_) cannam@127: { cannam@127: const problem_mpi_rdft *ego = (const problem_mpi_rdft *) ego_; cannam@127: R *I = ego->I; cannam@127: INT i, N; cannam@127: int my_pe; cannam@127: cannam@127: MPI_Comm_rank(ego->comm, &my_pe); cannam@127: N = ego->vn * XM(total_block)(ego->sz, IB, my_pe); cannam@127: for (i = 0; i < N; ++i) I[i] = K(0.0); cannam@127: } cannam@127: cannam@127: static const problem_adt padt = cannam@127: { cannam@127: PROBLEM_MPI_RDFT, cannam@127: hash, cannam@127: zero, cannam@127: print, cannam@127: destroy cannam@127: }; cannam@127: cannam@127: problem *XM(mkproblem_rdft)(const dtensor *sz, INT vn, cannam@127: R *I, R *O, cannam@127: MPI_Comm comm, cannam@127: const rdft_kind *kind, unsigned flags) cannam@127: { cannam@127: problem_mpi_rdft *ego; cannam@127: int i, rnk = sz->rnk; cannam@127: int n_pes; cannam@127: cannam@127: A(XM(dtensor_validp)(sz) && FINITE_RNK(sz->rnk)); cannam@127: MPI_Comm_size(comm, &n_pes); cannam@127: A(n_pes >= XM(num_blocks_total)(sz, IB) cannam@127: && n_pes >= XM(num_blocks_total)(sz, OB)); cannam@127: A(vn >= 0); cannam@127: cannam@127: #if defined(STRUCT_HACK_KR) cannam@127: ego = (problem_mpi_rdft *) X(mkproblem)(sizeof(problem_mpi_rdft) cannam@127: + sizeof(rdft_kind) cannam@127: * (rnk > 0 ? rnk - 1 : 0), &padt); cannam@127: #elif defined(STRUCT_HACK_C99) cannam@127: ego = (problem_mpi_rdft *) X(mkproblem)(sizeof(problem_mpi_rdft) cannam@127: + sizeof(rdft_kind) * rnk, &padt); cannam@127: #else cannam@127: ego = (problem_mpi_rdft *) X(mkproblem)(sizeof(problem_mpi_rdft), &padt); cannam@127: ego->kind = (rdft_kind *) MALLOC(sizeof(rdft_kind) * rnk, PROBLEMS); cannam@127: #endif cannam@127: cannam@127: /* enforce pointer equality if untainted pointers are equal */ cannam@127: if (UNTAINT(I) == UNTAINT(O)) cannam@127: I = O = JOIN_TAINT(I, O); cannam@127: cannam@127: ego->sz = XM(dtensor_canonical)(sz, 0); cannam@127: ego->vn = vn; cannam@127: ego->I = I; cannam@127: ego->O = O; cannam@127: for (i = 0; i< ego->sz->rnk; ++i) cannam@127: ego->kind[i] = kind[i]; cannam@127: cannam@127: /* canonicalize: replace TRANSPOSED_IN with TRANSPOSED_OUT by cannam@127: swapping the first two dimensions (for rnk > 1) */ cannam@127: if ((flags & TRANSPOSED_IN) && ego->sz->rnk > 1) { cannam@127: rdft_kind k = ego->kind[0]; cannam@127: ddim dim0 = ego->sz->dims[0]; cannam@127: ego->sz->dims[0] = ego->sz->dims[1]; cannam@127: ego->sz->dims[1] = dim0; cannam@127: ego->kind[0] = ego->kind[1]; cannam@127: ego->kind[1] = k; cannam@127: flags &= ~TRANSPOSED_IN; cannam@127: flags ^= TRANSPOSED_OUT; cannam@127: } cannam@127: ego->flags = flags; cannam@127: cannam@127: MPI_Comm_dup(comm, &ego->comm); cannam@127: cannam@127: return &(ego->super); cannam@127: } cannam@127: cannam@127: problem *XM(mkproblem_rdft_d)(dtensor *sz, INT vn, cannam@127: R *I, R *O, cannam@127: MPI_Comm comm, cannam@127: const rdft_kind *kind, unsigned flags) cannam@127: { cannam@127: problem *p = XM(mkproblem_rdft)(sz, vn, I, O, comm, kind, flags); cannam@127: XM(dtensor_destroy)(sz); cannam@127: return p; cannam@127: }