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: /* Complex RDFTs of rank == 1 when the vector length vn is >= # processes. cannam@95: In this case, we don't need to use a six-step type algorithm, and can cannam@95: instead transpose the RDFT dimension with the vector dimension to cannam@95: make the RDFT local. */ cannam@95: cannam@95: #include "mpi-rdft.h" cannam@95: #include "mpi-transpose.h" cannam@95: cannam@95: typedef struct { cannam@95: solver super; cannam@95: int preserve_input; /* preserve input even if DESTROY_INPUT was passed */ cannam@95: rearrangement rearrange; cannam@95: } S; cannam@95: cannam@95: typedef struct { cannam@95: plan_mpi_rdft super; cannam@95: cannam@95: plan *cldt_before, *cld, *cldt_after; cannam@95: int preserve_input; cannam@95: rearrangement rearrange; cannam@95: } P; cannam@95: cannam@95: static void apply(const plan *ego_, R *I, R *O) cannam@95: { cannam@95: const P *ego = (const P *) ego_; cannam@95: plan_rdft *cld, *cldt_before, *cldt_after; cannam@95: cannam@95: /* global transpose */ cannam@95: cldt_before = (plan_rdft *) ego->cldt_before; cannam@95: cldt_before->apply(ego->cldt_before, I, O); cannam@95: cannam@95: if (ego->preserve_input) I = O; cannam@95: cannam@95: /* 1d RDFT(s) */ cannam@95: cld = (plan_rdft *) ego->cld; cannam@95: cld->apply(ego->cld, O, I); cannam@95: cannam@95: /* global transpose */ cannam@95: cldt_after = (plan_rdft *) ego->cldt_after; cannam@95: cldt_after->apply(ego->cldt_after, I, O); cannam@95: } cannam@95: cannam@95: static int applicable(const S *ego, const problem *p_, cannam@95: const planner *plnr) cannam@95: { cannam@95: const problem_mpi_rdft *p = (const problem_mpi_rdft *) p_; cannam@95: int n_pes; cannam@95: MPI_Comm_size(p->comm, &n_pes); cannam@95: return (1 cannam@95: && p->sz->rnk == 1 cannam@95: && !(p->flags & ~RANK1_BIGVEC_ONLY) cannam@95: && (!ego->preserve_input || (!NO_DESTROY_INPUTP(plnr) cannam@95: && p->I != p->O)) cannam@95: cannam@95: #if 0 /* don't need this check since no other rank-1 rdft solver */ cannam@95: && (p->vn >= n_pes /* TODO: relax this, using more memory? */ cannam@95: || (p->flags & RANK1_BIGVEC_ONLY)) cannam@95: #endif cannam@95: cannam@95: && XM(rearrange_applicable)(ego->rearrange, cannam@95: p->sz->dims[0], p->vn, n_pes) cannam@95: cannam@95: && (!NO_SLOWP(plnr) /* slow if rdft-serial is applicable */ cannam@95: || !XM(rdft_serial_applicable)(p)) cannam@95: ); cannam@95: } cannam@95: cannam@95: static void awake(plan *ego_, enum wakefulness wakefulness) cannam@95: { cannam@95: P *ego = (P *) ego_; cannam@95: X(plan_awake)(ego->cldt_before, wakefulness); cannam@95: X(plan_awake)(ego->cld, wakefulness); cannam@95: X(plan_awake)(ego->cldt_after, wakefulness); cannam@95: } cannam@95: cannam@95: static void destroy(plan *ego_) cannam@95: { cannam@95: P *ego = (P *) ego_; cannam@95: X(plan_destroy_internal)(ego->cldt_after); cannam@95: X(plan_destroy_internal)(ego->cld); cannam@95: X(plan_destroy_internal)(ego->cldt_before); cannam@95: } cannam@95: cannam@95: static void print(const plan *ego_, printer *p) cannam@95: { cannam@95: const P *ego = (const P *) ego_; cannam@95: const char descrip[][16] = { "contig", "discontig", "square-after", cannam@95: "square-middle", "square-before" }; cannam@95: p->print(p, "(mpi-rdft-rank1-bigvec/%s%s %(%p%) %(%p%) %(%p%))", cannam@95: descrip[ego->rearrange], ego->preserve_input==2 ?"/p":"", cannam@95: ego->cldt_before, ego->cld, ego->cldt_after); cannam@95: } cannam@95: cannam@95: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr) cannam@95: { cannam@95: const S *ego = (const S *) ego_; cannam@95: const problem_mpi_rdft *p; cannam@95: P *pln; cannam@95: plan *cld = 0, *cldt_before = 0, *cldt_after = 0; cannam@95: R *I, *O; cannam@95: INT yblock, yb, nx, ny, vn; cannam@95: int my_pe, n_pes; cannam@95: static const plan_adt padt = { cannam@95: XM(rdft_solve), awake, print, destroy cannam@95: }; cannam@95: cannam@95: UNUSED(ego); cannam@95: cannam@95: if (!applicable(ego, p_, plnr)) cannam@95: return (plan *) 0; cannam@95: cannam@95: p = (const problem_mpi_rdft *) p_; cannam@95: cannam@95: MPI_Comm_rank(p->comm, &my_pe); cannam@95: MPI_Comm_size(p->comm, &n_pes); cannam@95: cannam@95: nx = p->sz->dims[0].n; cannam@95: if (!(ny = XM(rearrange_ny)(ego->rearrange, p->sz->dims[0],p->vn,n_pes))) cannam@95: return (plan *) 0; cannam@95: vn = p->vn / ny; cannam@95: A(ny * vn == p->vn); cannam@95: cannam@95: yblock = XM(default_block)(ny, n_pes); cannam@95: cldt_before = X(mkplan_d)(plnr, cannam@95: XM(mkproblem_transpose)( cannam@95: nx, ny, vn, cannam@95: I = p->I, O = p->O, cannam@95: p->sz->dims[0].b[IB], yblock, cannam@95: p->comm, 0)); cannam@95: if (XM(any_true)(!cldt_before, p->comm)) goto nada; cannam@95: if (ego->preserve_input || NO_DESTROY_INPUTP(plnr)) { I = O; } cannam@95: cannam@95: yb = XM(block)(ny, yblock, my_pe); cannam@95: cld = X(mkplan_d)(plnr, cannam@95: X(mkproblem_rdft_1_d)(X(mktensor_1d)(nx, vn, vn), cannam@95: X(mktensor_2d)(yb, vn*nx, vn*nx, cannam@95: vn, 1, 1), cannam@95: O, I, p->kind[0])); cannam@95: if (XM(any_true)(!cld, p->comm)) goto nada; cannam@95: cannam@95: cldt_after = X(mkplan_d)(plnr, cannam@95: XM(mkproblem_transpose)( cannam@95: ny, nx, vn, cannam@95: I, O, cannam@95: yblock, p->sz->dims[0].b[OB], cannam@95: p->comm, 0)); cannam@95: if (XM(any_true)(!cldt_after, p->comm)) goto nada; cannam@95: cannam@95: pln = MKPLAN_MPI_RDFT(P, &padt, apply); cannam@95: cannam@95: pln->cldt_before = cldt_before; cannam@95: pln->cld = cld; cannam@95: pln->cldt_after = cldt_after; cannam@95: pln->preserve_input = ego->preserve_input ? 2 : NO_DESTROY_INPUTP(plnr); cannam@95: pln->rearrange = ego->rearrange; cannam@95: cannam@95: X(ops_add)(&cldt_before->ops, &cld->ops, &pln->super.super.ops); cannam@95: X(ops_add2)(&cldt_after->ops, &pln->super.super.ops); cannam@95: cannam@95: return &(pln->super.super); cannam@95: cannam@95: nada: cannam@95: X(plan_destroy_internal)(cldt_after); cannam@95: X(plan_destroy_internal)(cld); cannam@95: X(plan_destroy_internal)(cldt_before); cannam@95: return (plan *) 0; cannam@95: } cannam@95: cannam@95: static solver *mksolver(rearrangement rearrange, int preserve_input) cannam@95: { cannam@95: static const solver_adt sadt = { PROBLEM_MPI_RDFT, mkplan, 0 }; cannam@95: S *slv = MKSOLVER(S, &sadt); cannam@95: slv->rearrange = rearrange; cannam@95: slv->preserve_input = preserve_input; cannam@95: return &(slv->super); cannam@95: } cannam@95: cannam@95: void XM(rdft_rank1_bigvec_register)(planner *p) cannam@95: { cannam@95: rearrangement rearrange; cannam@95: int preserve_input; cannam@95: FORALL_REARRANGE(rearrange) cannam@95: for (preserve_input = 0; preserve_input <= 1; ++preserve_input) cannam@95: REGISTER_SOLVER(p, mksolver(rearrange, preserve_input)); cannam@95: }