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: cannam@127: /* direct RDFT2 R2HC/HC2R solver, if we have a codelet */ cannam@127: cannam@127: #include "rdft.h" cannam@127: cannam@127: typedef struct { cannam@127: solver super; cannam@127: const kr2c_desc *desc; cannam@127: kr2c k; cannam@127: } S; cannam@127: cannam@127: typedef struct { cannam@127: plan_rdft2 super; cannam@127: cannam@127: stride rs, cs; cannam@127: INT vl; cannam@127: INT ivs, ovs; cannam@127: kr2c k; cannam@127: const S *slv; cannam@127: INT ilast; cannam@127: } P; cannam@127: cannam@127: static void apply(const plan *ego_, R *r0, R *r1, R *cr, R *ci) cannam@127: { cannam@127: const P *ego = (const P *) ego_; cannam@127: ASSERT_ALIGNED_DOUBLE; cannam@127: ego->k(r0, r1, cr, ci, cannam@127: ego->rs, ego->cs, ego->cs, cannam@127: ego->vl, ego->ivs, ego->ovs); cannam@127: } cannam@127: cannam@127: static void apply_r2hc(const plan *ego_, R *r0, R *r1, R *cr, R *ci) cannam@127: { cannam@127: const P *ego = (const P *) ego_; cannam@127: INT i, vl = ego->vl, ovs = ego->ovs; cannam@127: ASSERT_ALIGNED_DOUBLE; cannam@127: ego->k(r0, r1, cr, ci, cannam@127: ego->rs, ego->cs, ego->cs, cannam@127: vl, ego->ivs, ovs); cannam@127: for (i = 0; i < vl; ++i, ci += ovs) cannam@127: ci[0] = ci[ego->ilast] = 0; cannam@127: } cannam@127: cannam@127: static void destroy(plan *ego_) cannam@127: { cannam@127: P *ego = (P *) ego_; cannam@127: X(stride_destroy)(ego->rs); cannam@127: X(stride_destroy)(ego->cs); cannam@127: } cannam@127: cannam@127: static void print(const plan *ego_, printer *p) cannam@127: { cannam@127: const P *ego = (const P *) ego_; cannam@127: const S *s = ego->slv; cannam@127: cannam@127: p->print(p, "(rdft2-%s-direct-%D%v \"%s\")", cannam@127: X(rdft_kind_str)(s->desc->genus->kind), s->desc->n, cannam@127: ego->vl, s->desc->nam); cannam@127: } cannam@127: cannam@127: static int applicable(const solver *ego_, const problem *p_) cannam@127: { cannam@127: const S *ego = (const S *) ego_; cannam@127: const kr2c_desc *desc = ego->desc; cannam@127: const problem_rdft2 *p = (const problem_rdft2 *) p_; cannam@127: INT vl; cannam@127: INT ivs, ovs; cannam@127: cannam@127: return ( cannam@127: 1 cannam@127: && p->sz->rnk == 1 cannam@127: && p->vecsz->rnk <= 1 cannam@127: && p->sz->dims[0].n == desc->n cannam@127: && p->kind == desc->genus->kind cannam@127: cannam@127: /* check strides etc */ cannam@127: && X(tensor_tornk1)(p->vecsz, &vl, &ivs, &ovs) cannam@127: cannam@127: && (0 cannam@127: /* can operate out-of-place */ cannam@127: || p->r0 != p->cr cannam@127: cannam@127: /* cannam@127: * can compute one transform in-place, no matter cannam@127: * what the strides are. cannam@127: */ cannam@127: || p->vecsz->rnk == 0 cannam@127: cannam@127: /* can operate in-place as long as strides are the same */ cannam@127: || X(rdft2_inplace_strides)(p, RNK_MINFTY) cannam@127: ) cannam@127: ); cannam@127: } cannam@127: cannam@127: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr) cannam@127: { cannam@127: const S *ego = (const S *) ego_; cannam@127: P *pln; cannam@127: const problem_rdft2 *p; cannam@127: iodim *d; cannam@127: int r2hc_kindp; cannam@127: cannam@127: static const plan_adt padt = { cannam@127: X(rdft2_solve), X(null_awake), print, destroy cannam@127: }; cannam@127: cannam@127: UNUSED(plnr); cannam@127: cannam@127: if (!applicable(ego_, p_)) cannam@127: return (plan *)0; cannam@127: cannam@127: p = (const problem_rdft2 *) p_; cannam@127: cannam@127: r2hc_kindp = R2HC_KINDP(p->kind); cannam@127: A(r2hc_kindp || HC2R_KINDP(p->kind)); cannam@127: cannam@127: pln = MKPLAN_RDFT2(P, &padt, p->kind == R2HC ? apply_r2hc : apply); cannam@127: cannam@127: d = p->sz->dims; cannam@127: cannam@127: pln->k = ego->k; cannam@127: cannam@127: pln->rs = X(mkstride)(d->n, r2hc_kindp ? d->is : d->os); cannam@127: pln->cs = X(mkstride)(d->n, r2hc_kindp ? d->os : d->is); cannam@127: cannam@127: X(tensor_tornk1)(p->vecsz, &pln->vl, &pln->ivs, &pln->ovs); cannam@127: cannam@127: /* Nyquist freq., if any */ cannam@127: pln->ilast = (d->n % 2) ? 0 : (d->n/2) * d->os; cannam@127: cannam@127: pln->slv = ego; cannam@127: X(ops_zero)(&pln->super.super.ops); cannam@127: X(ops_madd2)(pln->vl / ego->desc->genus->vl, cannam@127: &ego->desc->ops, cannam@127: &pln->super.super.ops); cannam@127: if (p->kind == R2HC) cannam@127: pln->super.super.ops.other += 2 * pln->vl; /* + 2 stores */ cannam@127: cannam@127: pln->super.super.could_prune_now_p = 1; cannam@127: return &(pln->super.super); cannam@127: } cannam@127: cannam@127: /* constructor */ cannam@127: solver *X(mksolver_rdft2_direct)(kr2c k, const kr2c_desc *desc) cannam@127: { cannam@127: static const solver_adt sadt = { PROBLEM_RDFT2, mkplan, 0 }; cannam@127: S *slv = MKSOLVER(S, &sadt); cannam@127: slv->k = k; cannam@127: slv->desc = desc; cannam@127: return &(slv->super); cannam@127: }