From 5a422aba704c375a307a902bafe658342e209906 Mon Sep 17 00:00:00 2001 From: scuri Date: Fri, 17 Oct 2008 06:10:15 +0000 Subject: First commit - moving from LuaForge to SourceForge --- src/fftw3/rdft/hc2hc-buf.c | 245 +++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 245 insertions(+) create mode 100644 src/fftw3/rdft/hc2hc-buf.c (limited to 'src/fftw3/rdft/hc2hc-buf.c') diff --git a/src/fftw3/rdft/hc2hc-buf.c b/src/fftw3/rdft/hc2hc-buf.c new file mode 100644 index 0000000..75f761c --- /dev/null +++ b/src/fftw3/rdft/hc2hc-buf.c @@ -0,0 +1,245 @@ +/* + * Copyright (c) 2003 Matteo Frigo + * Copyright (c) 2003 Massachusetts Institute of Technology + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License as published by + * the Free Software Foundation; either version 2 of the License, or + * (at your option) any later version. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA + * + */ + +/* $Id: hc2hc-buf.c,v 1.1 2008/10/17 06:11:29 scuri Exp $ */ + +/* decimation in time Cooley-Tukey */ +#include "rdft.h" +#include "hc2hc.h" + +/* + Copy A -> B, where A and B are n0 x n1 complex matrices + such that the (i0, i1) element has index (i0 * s0 + i1 * s1). + The imaginary strides are of opposite signs to the real strides. +*/ +static void cpy(int n0, int n1, + const R *rA, const R *iA, int sa0, int sa1, + R *rB, R *iB, int sb0, int sb1) +{ + int i0, i1; + + for (i0 = 0; i0 < n0; ++i0) { + const R *pra, *pia; + R *prb, *pib; + pra = rA; rA += sa0; + pia = iA; iA -= sa0; + prb = rB; rB += sb0; + pib = iB; iB -= sb0; + + for (i1 = 0; i1 < n1; ++i1) { + R xr, xi; + xr = *pra; pra += sa1; + xi = *pia; pia -= sa1; + *prb = xr; prb += sb1; + *pib = xi; pib -= sb1; + } + } +} + +static const R *doit(khc2hc k, R *rA, R *iA, const R *W, int ios, int dist, + int r, int batchsz, R *buf, stride bufstride) +{ + cpy(r, batchsz, rA, iA, ios, dist, buf, buf + 2*batchsz*r-1, 1, r); + W = k(buf, buf + 2*batchsz*r-1, W, bufstride, 2*batchsz + 1, r); + cpy(r, batchsz, buf, buf + 2*batchsz*r-1, 1, r, rA, iA, ios, dist); + return W; +} + +#define BATCHSZ 4 /* FIXME: parametrize? */ + +static void apply_dit(const plan *ego_, R *I, R *O) +{ + const plan_hc2hc *ego = (const plan_hc2hc *) ego_; + + /* two-dimensional r x vl sub-transform: */ + { + plan_rdft *cld = (plan_rdft *) ego->cld; + cld->apply((plan *) cld, I, O); + } + + { + plan_rdft *cld0 = (plan_rdft *) ego->cld0; + plan_rdft *cldm = (plan_rdft *) ego->cldm; + int i, j, r = ego->r, m = ego->m, vl = ego->vl; + int os = ego->os, ovs = ego->ovs, ios = ego->iios; + R *buf; + + STACK_MALLOC(R *, buf, r * BATCHSZ * 2 * sizeof(R)); + + for (i = 0; i < vl; ++i, O += ovs) { + R *rA, *iA; + const R *W; + + cld0->apply((plan *) cld0, O, O); + + rA = O + os; iA = O + (r * m - 1) * os; + W = ego->W; + for (j = (m-1)/2; j >= BATCHSZ; j -= BATCHSZ) { + W = doit(ego->k, rA, iA, W, ios, os, r, BATCHSZ, buf, + ego->vs); + rA += os * (int)BATCHSZ; + iA -= os * (int)BATCHSZ; + } + /* do remaining j calls, if any */ + if (j > 0) + doit(ego->k, rA, iA, W, ios, os, r, j, buf, ego->vs); + + cldm->apply((plan *) cldm, O + os*(m/2), O + os*(m/2)); + } + + STACK_FREE(buf); + } +} + +static void apply_dif(const plan *ego_, R *I, R *O) +{ + const plan_hc2hc *ego = (const plan_hc2hc *) ego_; + R *I0 = I; + + { + plan_rdft *cld0 = (plan_rdft *) ego->cld0; + plan_rdft *cldm = (plan_rdft *) ego->cldm; + int i, j, r = ego->r, m = ego->m, vl = ego->vl; + int is = ego->is, ivs = ego->ivs, ios = ego->iios; + R *buf; + + STACK_MALLOC(R *, buf, r * BATCHSZ * 2 * sizeof(R)); + + for (i = 0; i < vl; ++i, I += ivs) { + R *rA, *iA; + const R *W; + + cld0->apply((plan *) cld0, I, I); + + rA = I + is; iA = I + (r * m - 1) * is; + W = ego->W; + for (j = (m-1)/2; j >= BATCHSZ; j -= BATCHSZ) { + W = doit(ego->k, rA, iA, W, ios, is, r, BATCHSZ, buf, + ego->vs); + rA += is * (int)BATCHSZ; + iA -= is * (int)BATCHSZ; + } + /* do remaining j calls, if any */ + if (j > 0) + doit(ego->k, rA, iA, W, ios, is, r, j, buf, ego->vs); + + cldm->apply((plan *) cldm, I + is*(m/2), I + is*(m/2)); + } + + STACK_FREE(buf); + } + + /* two-dimensional r x vl sub-transform: */ + { + plan_rdft *cld = (plan_rdft *) ego->cld; + cld->apply((plan *) cld, I0, O); + } +} + +static int applicable0(const solver_hc2hc *ego, const problem *p_, + const planner *plnr) +{ + if (X(rdft_hc2hc_applicable)(ego, p_)) { + const hc2hc_desc *e = ego->desc; + const problem_rdft *p = (const problem_rdft *) p_; + iodim *d = p->sz->dims; + int r = e->radix, m = d[0].n / e->radix; + return (1 + && (p->kind[0]==R2HC || p->I == p->O || DESTROY_INPUTP(plnr)) + /* check both batch size and remainder */ + && (m < BATCHSZ || + (e->genus->okp(e, 0, ((const R *)0)+2*BATCHSZ*r-1, 1,0, + 2*BATCHSZ + 1, r))) + && (m < BATCHSZ || + (e->genus->okp(e, 0, ((const R *)0) + + 2*(((m-1)/2) % BATCHSZ)*r-1, 1, 0, + 2*(((m-1)/2) % BATCHSZ) + 1, r))) + ); + } + return 0; +} + +static int applicable(const solver_hc2hc *ego, const problem *p_, + const planner *plnr) +{ + const problem_rdft *p; + + if (!applicable0(ego, p_, plnr)) return 0; + + p = (const problem_rdft *) p_; + + /* emulate fftw2 behavior */ + if (NO_VRECURSEP(plnr) && (p->vecsz->rnk > 0)) return 0; + + if (NO_UGLYP(plnr) && + X(ct_uglyp)(512, p->sz->dims[0].n, ego->desc->radix)) + return 0; + + return 1; +} + +static void finish(plan_hc2hc *ego) +{ + const hc2hc_desc *d = ego->slv->desc; + opcnt t; + + ego->iios = ego->m * (R2HC_KINDP(d->genus->kind) ? ego->os : ego->is); + ego->vs = X(mkstride)(ego->r, 1); + + X(ops_add)(&ego->cld0->ops, &ego->cldm->ops, &t); + X(ops_madd)(ego->vl, &t, &ego->cld->ops, &ego->super.super.ops); + ego->super.super.ops.other += 4 * ego->r * ((ego->m - 1)/2) * ego->vl; + X(ops_madd2)(ego->vl * ((ego->m - 1)/2) / d->genus->vl, &d->ops, + &ego->super.super.ops); +} + +static plan *mkplan_ditbuf(const solver *ego, const problem *p, planner *plnr) +{ + static const hc2hcadt adt = { + sizeof(plan_hc2hc), + X(rdft_mkcldrn_dit), finish, applicable, apply_dit + }; + return X(mkplan_rdft_hc2hc)((const solver_hc2hc *) ego, p, plnr, &adt); +} + +solver *X(mksolver_rdft_hc2hc_ditbuf)(khc2hc codelet, const hc2hc_desc *desc) +{ + static const solver_adt sadt = { mkplan_ditbuf }; + static const char name[] = "rdft-ditbuf"; + + return X(mksolver_rdft_hc2hc)(codelet, desc, name, &sadt); +} + +static plan *mkplan_difbuf(const solver *ego, const problem *p, planner *plnr) +{ + static const hc2hcadt adt = { + sizeof(plan_hc2hc), + X(rdft_mkcldrn_dif), finish, applicable, apply_dif + }; + return X(mkplan_rdft_hc2hc)((const solver_hc2hc *) ego, p, plnr, &adt); +} + +solver *X(mksolver_rdft_hc2hc_difbuf)(khc2hc codelet, const hc2hc_desc *desc) +{ + static const solver_adt sadt = { mkplan_difbuf }; + static const char name[] = "rdft-difbuf"; + + return X(mksolver_rdft_hc2hc)(codelet, desc, name, &sadt); +} -- cgit v1.2.3