Chris@10: /*
Chris@10:  * Copyright (c) 2003, 2007-11 Matteo Frigo
Chris@10:  * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
Chris@10:  *
Chris@10:  * This program is free software; you can redistribute it and/or modify
Chris@10:  * it under the terms of the GNU General Public License as published by
Chris@10:  * the Free Software Foundation; either version 2 of the License, or
Chris@10:  * (at your option) any later version.
Chris@10:  *
Chris@10:  * This program is distributed in the hope that it will be useful,
Chris@10:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@10:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
Chris@10:  * GNU General Public License for more details.
Chris@10:  *
Chris@10:  * You should have received a copy of the GNU General Public License
Chris@10:  * along with this program; if not, write to the Free Software
Chris@10:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
Chris@10:  *
Chris@10:  */
Chris@10: 
Chris@10: #include "dft.h"
Chris@10: 
Chris@10: typedef struct {
Chris@10:      solver super;
Chris@10: } S;
Chris@10: 
Chris@10: typedef struct {
Chris@10:      plan_dft super;
Chris@10:      INT n;     /* problem size */
Chris@10:      INT nb;    /* size of convolution */
Chris@10:      R *w;      /* lambda k . exp(2*pi*i*k^2/(2*n)) */
Chris@10:      R *W;      /* DFT(w) */
Chris@10:      plan *cldf;
Chris@10:      INT is, os;
Chris@10: } P;
Chris@10: 
Chris@10: static void bluestein_sequence(enum wakefulness wakefulness, INT n, R *w)
Chris@10: {
Chris@10:      INT k, ksq, n2 = 2 * n;
Chris@10:      triggen *t = X(mktriggen)(wakefulness, n2);
Chris@10: 
Chris@10:      ksq = 0;
Chris@10:      for (k = 0; k < n; ++k) {
Chris@10: 	  t->cexp(t, ksq, w+2*k);
Chris@10:           /* careful with overflow */
Chris@10:           ksq += 2*k + 1; while (ksq > n2) ksq -= n2;
Chris@10:      }
Chris@10: 
Chris@10:      X(triggen_destroy)(t);
Chris@10: }
Chris@10: 
Chris@10: static void mktwiddle(enum wakefulness wakefulness, P *p)
Chris@10: {
Chris@10:      INT i;
Chris@10:      INT n = p->n, nb = p->nb;
Chris@10:      R *w, *W;
Chris@10:      E nbf = (E)nb;
Chris@10: 
Chris@10:      p->w = w = (R *) MALLOC(2 * n * sizeof(R), TWIDDLES);
Chris@10:      p->W = W = (R *) MALLOC(2 * nb * sizeof(R), TWIDDLES);
Chris@10: 
Chris@10:      bluestein_sequence(wakefulness, n, w);
Chris@10: 
Chris@10:      for (i = 0; i < nb; ++i)
Chris@10:           W[2*i] = W[2*i+1] = K(0.0);
Chris@10: 
Chris@10:      W[0] = w[0] / nbf;
Chris@10:      W[1] = w[1] / nbf;
Chris@10: 
Chris@10:      for (i = 1; i < n; ++i) {
Chris@10:           W[2*i] = W[2*(nb-i)] = w[2*i] / nbf;
Chris@10:           W[2*i+1] = W[2*(nb-i)+1] = w[2*i+1] / nbf;
Chris@10:      }
Chris@10: 
Chris@10:      {
Chris@10:           plan_dft *cldf = (plan_dft *)p->cldf;
Chris@10: 	  /* cldf must be awake */
Chris@10:           cldf->apply(p->cldf, W, W+1, W, W+1);
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: static void apply(const plan *ego_, R *ri, R *ii, R *ro, R *io)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      INT i, n = ego->n, nb = ego->nb, is = ego->is, os = ego->os;
Chris@10:      R *w = ego->w, *W = ego->W;
Chris@10:      R *b = (R *) MALLOC(2 * nb * sizeof(R), BUFFERS);
Chris@10: 
Chris@10:      /* multiply input by conjugate bluestein sequence */
Chris@10:      for (i = 0; i < n; ++i) {
Chris@10: 	  E xr = ri[i*is], xi = ii[i*is];
Chris@10:           E wr = w[2*i], wi = w[2*i+1];
Chris@10:           b[2*i] = xr * wr + xi * wi;
Chris@10:           b[2*i+1] = xi * wr - xr * wi;
Chris@10:      }
Chris@10: 
Chris@10:      for (; i < nb; ++i) b[2*i] = b[2*i+1] = K(0.0);
Chris@10: 
Chris@10:      /* convolution: FFT */
Chris@10:      {
Chris@10:           plan_dft *cldf = (plan_dft *)ego->cldf;
Chris@10:           cldf->apply(ego->cldf, b, b+1, b, b+1);
Chris@10:      }
Chris@10: 
Chris@10:      /* convolution: pointwise multiplication */
Chris@10:      for (i = 0; i < nb; ++i) {
Chris@10: 	  E xr = b[2*i], xi = b[2*i+1];
Chris@10:           E wr = W[2*i], wi = W[2*i+1];
Chris@10:           b[2*i] = xi * wr + xr * wi;
Chris@10:           b[2*i+1] = xr * wr - xi * wi;
Chris@10:      }
Chris@10: 
Chris@10:      /* convolution: IFFT by FFT with real/imag input/output swapped */
Chris@10:      {
Chris@10:           plan_dft *cldf = (plan_dft *)ego->cldf;
Chris@10:           cldf->apply(ego->cldf, b, b+1, b, b+1);
Chris@10:      }
Chris@10: 
Chris@10:      /* multiply output by conjugate bluestein sequence */
Chris@10:      for (i = 0; i < n; ++i) {
Chris@10: 	  E xi = b[2*i], xr = b[2*i+1];
Chris@10:           E wr = w[2*i], wi = w[2*i+1];
Chris@10:           ro[i*os] = xr * wr + xi * wi;
Chris@10:           io[i*os] = xi * wr - xr * wi;
Chris@10:      }
Chris@10: 
Chris@10:      X(ifree)(b);	  
Chris@10: }
Chris@10: 
Chris@10: static void awake(plan *ego_, enum wakefulness wakefulness)
Chris@10: {
Chris@10:      P *ego = (P *) ego_;
Chris@10: 
Chris@10:      X(plan_awake)(ego->cldf, wakefulness);
Chris@10: 
Chris@10:      switch (wakefulness) {
Chris@10: 	 case SLEEPY:
Chris@10: 	      X(ifree0)(ego->w); ego->w = 0;
Chris@10: 	      X(ifree0)(ego->W); ego->W = 0;
Chris@10: 	      break;
Chris@10: 	 default:
Chris@10: 	      A(!ego->w);
Chris@10: 	      mktwiddle(wakefulness, ego);
Chris@10: 	      break;
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: static int applicable(const solver *ego, const problem *p_, 
Chris@10: 		      const planner *plnr)
Chris@10: {
Chris@10:      const problem_dft *p = (const problem_dft *) p_;
Chris@10:      UNUSED(ego);
Chris@10:      return (1
Chris@10: 	     && p->sz->rnk == 1
Chris@10: 	     && p->vecsz->rnk == 0
Chris@10: 	     /* FIXME: allow other sizes */
Chris@10: 	     && X(is_prime)(p->sz->dims[0].n)
Chris@10: 
Chris@10: 	     /* FIXME: avoid infinite recursion of bluestein with itself.
Chris@10: 		This works because all factors in child problems are 2, 3, 5 */
Chris@10: 	     && p->sz->dims[0].n > 16
Chris@10: 
Chris@10: 	     && CIMPLIES(NO_SLOWP(plnr), p->sz->dims[0].n > BLUESTEIN_MAX_SLOW)
Chris@10: 	  );
Chris@10: }
Chris@10: 
Chris@10: static void destroy(plan *ego_)
Chris@10: {
Chris@10:      P *ego = (P *) ego_;
Chris@10:      X(plan_destroy_internal)(ego->cldf);
Chris@10: }
Chris@10: 
Chris@10: static void print(const plan *ego_, printer *p)
Chris@10: {
Chris@10:      const P *ego = (const P *)ego_;
Chris@10:      p->print(p, "(dft-bluestein-%D/%D%(%p%))",
Chris@10:               ego->n, ego->nb, ego->cldf);
Chris@10: }
Chris@10: 
Chris@10: static INT choose_transform_size(INT minsz)
Chris@10: {
Chris@10:      while (!X(factors_into_small_primes)(minsz))
Chris@10: 	  ++minsz;
Chris@10:      return minsz;
Chris@10: }
Chris@10: 
Chris@10: static plan *mkplan(const solver *ego, const problem *p_, planner *plnr)
Chris@10: {
Chris@10:      const problem_dft *p = (const problem_dft *) p_;
Chris@10:      P *pln;
Chris@10:      INT n, nb;
Chris@10:      plan *cldf = 0;
Chris@10:      R *buf = (R *) 0;
Chris@10: 
Chris@10:      static const plan_adt padt = {
Chris@10: 	  X(dft_solve), awake, print, destroy
Chris@10:      };
Chris@10: 
Chris@10:      if (!applicable(ego, p_, plnr))
Chris@10: 	  return (plan *) 0;
Chris@10: 
Chris@10:      n = p->sz->dims[0].n;
Chris@10:      nb = choose_transform_size(2 * n - 1);
Chris@10:      buf = (R *) MALLOC(2 * nb * sizeof(R), BUFFERS);
Chris@10: 
Chris@10:      cldf = X(mkplan_f_d)(plnr, 
Chris@10: 			  X(mkproblem_dft_d)(X(mktensor_1d)(nb, 2, 2),
Chris@10: 					     X(mktensor_1d)(1, 0, 0),
Chris@10: 					     buf, buf+1, 
Chris@10: 					     buf, buf+1),
Chris@10: 			  NO_SLOW, 0, 0);
Chris@10:      if (!cldf) goto nada;
Chris@10: 
Chris@10:      X(ifree)(buf);
Chris@10: 
Chris@10:      pln = MKPLAN_DFT(P, &padt, apply);
Chris@10: 
Chris@10:      pln->n = n;
Chris@10:      pln->nb = nb;
Chris@10:      pln->w = 0;
Chris@10:      pln->W = 0;
Chris@10:      pln->cldf = cldf;
Chris@10:      pln->is = p->sz->dims[0].is;
Chris@10:      pln->os = p->sz->dims[0].os;
Chris@10: 
Chris@10:      X(ops_add)(&cldf->ops, &cldf->ops, &pln->super.super.ops);
Chris@10:      pln->super.super.ops.add += 4 * n + 2 * nb;
Chris@10:      pln->super.super.ops.mul += 8 * n + 4 * nb;
Chris@10:      pln->super.super.ops.other += 6 * (n + nb);
Chris@10: 
Chris@10:      return &(pln->super.super);
Chris@10: 
Chris@10:  nada:
Chris@10:      X(ifree0)(buf);
Chris@10:      X(plan_destroy_internal)(cldf);
Chris@10:      return (plan *)0;
Chris@10: }
Chris@10: 
Chris@10: 
Chris@10: static solver *mksolver(void)
Chris@10: {
Chris@10:      static const solver_adt sadt = { PROBLEM_DFT, mkplan, 0 };
Chris@10:      S *slv = MKSOLVER(S, &sadt);
Chris@10:      return &(slv->super);
Chris@10: }
Chris@10: 
Chris@10: void X(dft_bluestein_register)(planner *p)
Chris@10: {
Chris@10:      REGISTER_SOLVER(p, mksolver());
Chris@10: }