cannam@95: /* cannam@95: * Copyright (c) 2003, 2007-11 Matteo Frigo cannam@95: * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology cannam@95: * cannam@95: * This program is free software; you can redistribute it and/or modify cannam@95: * it under the terms of the GNU General Public License as published by cannam@95: * the Free Software Foundation; either version 2 of the License, or cannam@95: * (at your option) any later version. cannam@95: * cannam@95: * This program is distributed in the hope that it will be useful, cannam@95: * but WITHOUT ANY WARRANTY; without even the implied warranty of cannam@95: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the cannam@95: * GNU General Public License for more details. cannam@95: * cannam@95: * You should have received a copy of the GNU General Public License cannam@95: * along with this program; if not, write to the Free Software cannam@95: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA cannam@95: * cannam@95: */ cannam@95: cannam@95: #include "mpi-rdft2.h" cannam@95: cannam@95: static void destroy(problem *ego_) cannam@95: { cannam@95: problem_mpi_rdft2 *ego = (problem_mpi_rdft2 *) ego_; cannam@95: XM(dtensor_destroy)(ego->sz); cannam@95: MPI_Comm_free(&ego->comm); cannam@95: X(ifree)(ego_); cannam@95: } cannam@95: cannam@95: static void hash(const problem *p_, md5 *m) cannam@95: { cannam@95: const problem_mpi_rdft2 *p = (const problem_mpi_rdft2 *) p_; cannam@95: int i; cannam@95: X(md5puts)(m, "mpi-rdft2"); cannam@95: X(md5int)(m, p->I == p->O); cannam@95: /* don't include alignment -- may differ between processes cannam@95: X(md5int)(m, X(alignment_of)(p->I)); cannam@95: X(md5int)(m, X(alignment_of)(p->O)); cannam@95: ... note that applicability of MPI plans does not depend cannam@95: on alignment (although optimality may, in principle). */ cannam@95: XM(dtensor_md5)(m, p->sz); cannam@95: X(md5INT)(m, p->vn); cannam@95: X(md5int)(m, p->kind); cannam@95: X(md5int)(m, p->flags); cannam@95: MPI_Comm_size(p->comm, &i); X(md5int)(m, i); cannam@95: A(XM(md5_equal)(*m, p->comm)); cannam@95: } cannam@95: cannam@95: static void print(const problem *ego_, printer *p) cannam@95: { cannam@95: const problem_mpi_rdft2 *ego = (const problem_mpi_rdft2 *) ego_; cannam@95: int i; cannam@95: p->print(p, "(mpi-rdft2 %d %d %d ", cannam@95: ego->I == ego->O, cannam@95: X(alignment_of)(ego->I), cannam@95: X(alignment_of)(ego->O)); cannam@95: XM(dtensor_print)(ego->sz, p); cannam@95: p->print(p, " %D %d %d", ego->vn, (int) ego->kind, ego->flags); cannam@95: MPI_Comm_size(ego->comm, &i); p->print(p, " %d)", i); cannam@95: } cannam@95: cannam@95: static void zero(const problem *ego_) cannam@95: { cannam@95: const problem_mpi_rdft2 *ego = (const problem_mpi_rdft2 *) ego_; cannam@95: R *I = ego->I; cannam@95: dtensor *sz; cannam@95: INT i, N; cannam@95: int my_pe; cannam@95: cannam@95: sz = XM(dtensor_copy)(ego->sz); cannam@95: sz->dims[sz->rnk - 1].n = sz->dims[sz->rnk - 1].n / 2 + 1; cannam@95: MPI_Comm_rank(ego->comm, &my_pe); cannam@95: N = 2 * ego->vn * XM(total_block)(sz, IB, my_pe); cannam@95: XM(dtensor_destroy)(sz); cannam@95: for (i = 0; i < N; ++i) I[i] = K(0.0); cannam@95: } cannam@95: cannam@95: static const problem_adt padt = cannam@95: { cannam@95: PROBLEM_MPI_RDFT2, cannam@95: hash, cannam@95: zero, cannam@95: print, cannam@95: destroy cannam@95: }; cannam@95: cannam@95: problem *XM(mkproblem_rdft2)(const dtensor *sz, INT vn, cannam@95: R *I, R *O, cannam@95: MPI_Comm comm, cannam@95: rdft_kind kind, cannam@95: unsigned flags) cannam@95: { cannam@95: problem_mpi_rdft2 *ego = cannam@95: (problem_mpi_rdft2 *)X(mkproblem)(sizeof(problem_mpi_rdft2), &padt); cannam@95: int n_pes; cannam@95: cannam@95: A(XM(dtensor_validp)(sz) && FINITE_RNK(sz->rnk) && sz->rnk > 1); cannam@95: MPI_Comm_size(comm, &n_pes); cannam@95: A(vn >= 0); cannam@95: A(kind == R2HC || kind == HC2R); cannam@95: cannam@95: /* enforce pointer equality if untainted pointers are equal */ cannam@95: if (UNTAINT(I) == UNTAINT(O)) cannam@95: I = O = JOIN_TAINT(I, O); cannam@95: cannam@95: ego->sz = XM(dtensor_canonical)(sz, 0); cannam@95: #ifdef FFTW_DEBUG cannam@95: ego->sz->dims[sz->rnk - 1].n = sz->dims[sz->rnk - 1].n / 2 + 1; cannam@95: A(n_pes >= XM(num_blocks_total)(ego->sz, IB) cannam@95: && n_pes >= XM(num_blocks_total)(ego->sz, OB)); cannam@95: ego->sz->dims[sz->rnk - 1].n = sz->dims[sz->rnk - 1].n; cannam@95: #endif cannam@95: cannam@95: ego->vn = vn; cannam@95: ego->I = I; cannam@95: ego->O = O; cannam@95: ego->kind = kind; cannam@95: cannam@95: /* We only support TRANSPOSED_OUT for r2c and TRANSPOSED_IN for cannam@95: c2r transforms. */ cannam@95: cannam@95: ego->flags = flags; cannam@95: cannam@95: MPI_Comm_dup(comm, &ego->comm); cannam@95: cannam@95: return &(ego->super); cannam@95: } cannam@95: cannam@95: problem *XM(mkproblem_rdft2_d)(dtensor *sz, INT vn, cannam@95: R *I, R *O, cannam@95: MPI_Comm comm, cannam@95: rdft_kind kind, cannam@95: unsigned flags) cannam@95: { cannam@95: problem *p = XM(mkproblem_rdft2)(sz, vn, I, O, comm, kind, flags); cannam@95: XM(dtensor_destroy)(sz); cannam@95: return p; cannam@95: }