diff options
author | scuri <scuri> | 2009-08-20 12:35:06 +0000 |
---|---|---|
committer | scuri <scuri> | 2009-08-20 12:35:06 +0000 |
commit | 5d735255ddd3cb2f547abd3d03969af3fb7eb04e (patch) | |
tree | 8fb66510bc625bb1b08ccb41f1b83fb0f7cb8f19 /src/fftw3/reodft/redft00e-r2hc.c | |
parent | 35733b87eed86e5228f12fa10c98a3d9d22a6073 (diff) |
*** empty log message ***
Diffstat (limited to 'src/fftw3/reodft/redft00e-r2hc.c')
-rw-r--r-- | src/fftw3/reodft/redft00e-r2hc.c | 216 |
1 files changed, 0 insertions, 216 deletions
diff --git a/src/fftw3/reodft/redft00e-r2hc.c b/src/fftw3/reodft/redft00e-r2hc.c deleted file mode 100644 index 0cd742f..0000000 --- a/src/fftw3/reodft/redft00e-r2hc.c +++ /dev/null @@ -1,216 +0,0 @@ -/* - * 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: redft00e-r2hc.c,v 1.1 2008/10/17 06:13:18 scuri Exp $ */ - -/* Do a REDFT00 problem via an R2HC problem, with some pre/post-processing. - - This code uses the trick from FFTPACK, also documented in a similar - form by Numerical Recipes. Unfortunately, this algorithm seems to - have intrinsic numerical problems (similar to those in - reodft11e-r2hc.c), possibly due to the fact that it multiplies its - input by a cosine, causing a loss of precision near the zero. For - transforms of 16k points, it has already lost three or four decimal - places of accuracy, which we deem unacceptable. - - So, we have abandoned this algorithm in favor of the one in - redft00-r2hc-pad.c, which unfortunately sacrifices 30-50% in speed. - The only other alternative in the literature that does not have - similar numerical difficulties seems to be the direct adaptation of - the Cooley-Tukey decomposition for symmetric data, but this would - require a whole new set of codelets and it's not clear that it's - worth it at this point. */ - -#include "reodft.h" - -typedef struct { - solver super; -} S; - -typedef struct { - plan_rdft super; - plan *cld; - twid *td; - int is, os; - int n; - int vl; - int ivs, ovs; -} P; - -static void apply(const plan *ego_, R *I, R *O) -{ - const P *ego = (const P *) ego_; - int is = ego->is, os = ego->os; - int i, n = ego->n; - int iv, vl = ego->vl; - int ivs = ego->ivs, ovs = ego->ovs; - R *W = ego->td->W; - R *buf; - E csum; - - buf = (R *) MALLOC(sizeof(R) * n, BUFFERS); - - for (iv = 0; iv < vl; ++iv, I += ivs, O += ovs) { - buf[0] = I[0] + I[is * n]; - csum = I[0] - I[is * n]; - for (i = 1; i < n - i; ++i) { - E a, b, apb, amb; - a = I[is * i]; - b = I[is * (n - i)]; - csum += W[2*i] * (amb = K(2.0)*(a - b)); - amb = W[2*i+1] * amb; - apb = (a + b); - buf[i] = apb - amb; - buf[n - i] = apb + amb; - } - if (i == n - i) { - buf[i] = K(2.0) * I[is * i]; - } - - { - plan_rdft *cld = (plan_rdft *) ego->cld; - cld->apply((plan *) cld, buf, buf); - } - - /* FIXME: use recursive/cascade summation for better stability? */ - O[0] = buf[0]; - O[os] = csum; - for (i = 1; i + i < n; ++i) { - int k = i + i; - O[os * k] = buf[i]; - O[os * (k + 1)] = O[os * (k - 1)] - buf[n - i]; - } - if (i + i == n) { - O[os * n] = buf[i]; - } - } - - X(ifree)(buf); -} - -static void awake(plan *ego_, int flg) -{ - P *ego = (P *) ego_; - static const tw_instr redft00e_tw[] = { - { TW_COS, 0, 1 }, - { TW_SIN, 0, 1 }, - { TW_NEXT, 1, 0 } - }; - - AWAKE(ego->cld, flg); - X(twiddle_awake)(flg, &ego->td, redft00e_tw, 2*ego->n, 1, (ego->n+1)/2); -} - -static void destroy(plan *ego_) -{ - P *ego = (P *) ego_; - X(plan_destroy_internal)(ego->cld); -} - -static void print(const plan *ego_, printer *p) -{ - const P *ego = (const P *) ego_; - p->print(p, "(redft00e-r2hc-%d%v%(%p%))", ego->n + 1, ego->vl, ego->cld); -} - -static int applicable0(const solver *ego_, const problem *p_) -{ - UNUSED(ego_); - if (RDFTP(p_)) { - const problem_rdft *p = (const problem_rdft *) p_; - return (1 - && p->sz->rnk == 1 - && p->vecsz->rnk <= 1 - && p->kind[0] == REDFT00 - && p->sz->dims[0].n > 1 /* n == 1 is not well-defined */ - ); - } - - return 0; -} - -static int applicable(const solver *ego, const problem *p, const planner *plnr) -{ - return (!NO_UGLYP(plnr) && applicable0(ego, p)); -} - -static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr) -{ - P *pln; - const problem_rdft *p; - plan *cld; - R *buf; - int n; - opcnt ops; - - static const plan_adt padt = { - X(rdft_solve), awake, print, destroy - }; - - if (!applicable(ego_, p_, plnr)) - return (plan *)0; - - p = (const problem_rdft *) p_; - - n = p->sz->dims[0].n - 1; - A(n > 0); - buf = (R *) MALLOC(sizeof(R) * n, BUFFERS); - - cld = X(mkplan_d)(plnr, X(mkproblem_rdft_1_d)(X(mktensor_1d)(n, 1, 1), - X(mktensor_0d)(), - buf, buf, R2HC)); - X(ifree)(buf); - if (!cld) - return (plan *)0; - - pln = MKPLAN_RDFT(P, &padt, apply); - - pln->n = n; - pln->is = p->sz->dims[0].is; - pln->os = p->sz->dims[0].os; - pln->cld = cld; - pln->td = 0; - - X(tensor_tornk1)(p->vecsz, &pln->vl, &pln->ivs, &pln->ovs); - - X(ops_zero)(&ops); - ops.other = 8 + (n-1)/2 * 11 + (1 - n % 2) * 5; - ops.add = 2 + (n-1)/2 * 5; - ops.mul = (n-1)/2 * 3 + (1 - n % 2) * 1; - - X(ops_zero)(&pln->super.super.ops); - X(ops_madd2)(pln->vl, &ops, &pln->super.super.ops); - X(ops_madd2)(pln->vl, &cld->ops, &pln->super.super.ops); - - return &(pln->super.super); -} - -/* constructor */ -static solver *mksolver(void) -{ - static const solver_adt sadt = { mkplan }; - S *slv = MKSOLVER(S, &sadt); - return &(slv->super); -} - -void X(redft00e_r2hc_register)(planner *p) -{ - REGISTER_SOLVER(p, mksolver()); -} |