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: #include "dft.h"
cannam@95: 
cannam@95: typedef struct {
cannam@95:      solver super;
cannam@95: } S;
cannam@95: 
cannam@95: typedef struct {
cannam@95:      plan_dft super;
cannam@95:      INT n;     /* problem size */
cannam@95:      INT nb;    /* size of convolution */
cannam@95:      R *w;      /* lambda k . exp(2*pi*i*k^2/(2*n)) */
cannam@95:      R *W;      /* DFT(w) */
cannam@95:      plan *cldf;
cannam@95:      INT is, os;
cannam@95: } P;
cannam@95: 
cannam@95: static void bluestein_sequence(enum wakefulness wakefulness, INT n, R *w)
cannam@95: {
cannam@95:      INT k, ksq, n2 = 2 * n;
cannam@95:      triggen *t = X(mktriggen)(wakefulness, n2);
cannam@95: 
cannam@95:      ksq = 0;
cannam@95:      for (k = 0; k < n; ++k) {
cannam@95: 	  t->cexp(t, ksq, w+2*k);
cannam@95:           /* careful with overflow */
cannam@95:           ksq += 2*k + 1; while (ksq > n2) ksq -= n2;
cannam@95:      }
cannam@95: 
cannam@95:      X(triggen_destroy)(t);
cannam@95: }
cannam@95: 
cannam@95: static void mktwiddle(enum wakefulness wakefulness, P *p)
cannam@95: {
cannam@95:      INT i;
cannam@95:      INT n = p->n, nb = p->nb;
cannam@95:      R *w, *W;
cannam@95:      E nbf = (E)nb;
cannam@95: 
cannam@95:      p->w = w = (R *) MALLOC(2 * n * sizeof(R), TWIDDLES);
cannam@95:      p->W = W = (R *) MALLOC(2 * nb * sizeof(R), TWIDDLES);
cannam@95: 
cannam@95:      bluestein_sequence(wakefulness, n, w);
cannam@95: 
cannam@95:      for (i = 0; i < nb; ++i)
cannam@95:           W[2*i] = W[2*i+1] = K(0.0);
cannam@95: 
cannam@95:      W[0] = w[0] / nbf;
cannam@95:      W[1] = w[1] / nbf;
cannam@95: 
cannam@95:      for (i = 1; i < n; ++i) {
cannam@95:           W[2*i] = W[2*(nb-i)] = w[2*i] / nbf;
cannam@95:           W[2*i+1] = W[2*(nb-i)+1] = w[2*i+1] / nbf;
cannam@95:      }
cannam@95: 
cannam@95:      {
cannam@95:           plan_dft *cldf = (plan_dft *)p->cldf;
cannam@95: 	  /* cldf must be awake */
cannam@95:           cldf->apply(p->cldf, W, W+1, W, W+1);
cannam@95:      }
cannam@95: }
cannam@95: 
cannam@95: static void apply(const plan *ego_, R *ri, R *ii, R *ro, R *io)
cannam@95: {
cannam@95:      const P *ego = (const P *) ego_;
cannam@95:      INT i, n = ego->n, nb = ego->nb, is = ego->is, os = ego->os;
cannam@95:      R *w = ego->w, *W = ego->W;
cannam@95:      R *b = (R *) MALLOC(2 * nb * sizeof(R), BUFFERS);
cannam@95: 
cannam@95:      /* multiply input by conjugate bluestein sequence */
cannam@95:      for (i = 0; i < n; ++i) {
cannam@95: 	  E xr = ri[i*is], xi = ii[i*is];
cannam@95:           E wr = w[2*i], wi = w[2*i+1];
cannam@95:           b[2*i] = xr * wr + xi * wi;
cannam@95:           b[2*i+1] = xi * wr - xr * wi;
cannam@95:      }
cannam@95: 
cannam@95:      for (; i < nb; ++i) b[2*i] = b[2*i+1] = K(0.0);
cannam@95: 
cannam@95:      /* convolution: FFT */
cannam@95:      {
cannam@95:           plan_dft *cldf = (plan_dft *)ego->cldf;
cannam@95:           cldf->apply(ego->cldf, b, b+1, b, b+1);
cannam@95:      }
cannam@95: 
cannam@95:      /* convolution: pointwise multiplication */
cannam@95:      for (i = 0; i < nb; ++i) {
cannam@95: 	  E xr = b[2*i], xi = b[2*i+1];
cannam@95:           E wr = W[2*i], wi = W[2*i+1];
cannam@95:           b[2*i] = xi * wr + xr * wi;
cannam@95:           b[2*i+1] = xr * wr - xi * wi;
cannam@95:      }
cannam@95: 
cannam@95:      /* convolution: IFFT by FFT with real/imag input/output swapped */
cannam@95:      {
cannam@95:           plan_dft *cldf = (plan_dft *)ego->cldf;
cannam@95:           cldf->apply(ego->cldf, b, b+1, b, b+1);
cannam@95:      }
cannam@95: 
cannam@95:      /* multiply output by conjugate bluestein sequence */
cannam@95:      for (i = 0; i < n; ++i) {
cannam@95: 	  E xi = b[2*i], xr = b[2*i+1];
cannam@95:           E wr = w[2*i], wi = w[2*i+1];
cannam@95:           ro[i*os] = xr * wr + xi * wi;
cannam@95:           io[i*os] = xi * wr - xr * wi;
cannam@95:      }
cannam@95: 
cannam@95:      X(ifree)(b);	  
cannam@95: }
cannam@95: 
cannam@95: static void awake(plan *ego_, enum wakefulness wakefulness)
cannam@95: {
cannam@95:      P *ego = (P *) ego_;
cannam@95: 
cannam@95:      X(plan_awake)(ego->cldf, wakefulness);
cannam@95: 
cannam@95:      switch (wakefulness) {
cannam@95: 	 case SLEEPY:
cannam@95: 	      X(ifree0)(ego->w); ego->w = 0;
cannam@95: 	      X(ifree0)(ego->W); ego->W = 0;
cannam@95: 	      break;
cannam@95: 	 default:
cannam@95: 	      A(!ego->w);
cannam@95: 	      mktwiddle(wakefulness, ego);
cannam@95: 	      break;
cannam@95:      }
cannam@95: }
cannam@95: 
cannam@95: static int applicable(const solver *ego, const problem *p_, 
cannam@95: 		      const planner *plnr)
cannam@95: {
cannam@95:      const problem_dft *p = (const problem_dft *) p_;
cannam@95:      UNUSED(ego);
cannam@95:      return (1
cannam@95: 	     && p->sz->rnk == 1
cannam@95: 	     && p->vecsz->rnk == 0
cannam@95: 	     /* FIXME: allow other sizes */
cannam@95: 	     && X(is_prime)(p->sz->dims[0].n)
cannam@95: 
cannam@95: 	     /* FIXME: avoid infinite recursion of bluestein with itself.
cannam@95: 		This works because all factors in child problems are 2, 3, 5 */
cannam@95: 	     && p->sz->dims[0].n > 16
cannam@95: 
cannam@95: 	     && CIMPLIES(NO_SLOWP(plnr), p->sz->dims[0].n > BLUESTEIN_MAX_SLOW)
cannam@95: 	  );
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->cldf);
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:      p->print(p, "(dft-bluestein-%D/%D%(%p%))",
cannam@95:               ego->n, ego->nb, ego->cldf);
cannam@95: }
cannam@95: 
cannam@95: static INT choose_transform_size(INT minsz)
cannam@95: {
cannam@95:      while (!X(factors_into_small_primes)(minsz))
cannam@95: 	  ++minsz;
cannam@95:      return minsz;
cannam@95: }
cannam@95: 
cannam@95: static plan *mkplan(const solver *ego, const problem *p_, planner *plnr)
cannam@95: {
cannam@95:      const problem_dft *p = (const problem_dft *) p_;
cannam@95:      P *pln;
cannam@95:      INT n, nb;
cannam@95:      plan *cldf = 0;
cannam@95:      R *buf = (R *) 0;
cannam@95: 
cannam@95:      static const plan_adt padt = {
cannam@95: 	  X(dft_solve), awake, print, destroy
cannam@95:      };
cannam@95: 
cannam@95:      if (!applicable(ego, p_, plnr))
cannam@95: 	  return (plan *) 0;
cannam@95: 
cannam@95:      n = p->sz->dims[0].n;
cannam@95:      nb = choose_transform_size(2 * n - 1);
cannam@95:      buf = (R *) MALLOC(2 * nb * sizeof(R), BUFFERS);
cannam@95: 
cannam@95:      cldf = X(mkplan_f_d)(plnr, 
cannam@95: 			  X(mkproblem_dft_d)(X(mktensor_1d)(nb, 2, 2),
cannam@95: 					     X(mktensor_1d)(1, 0, 0),
cannam@95: 					     buf, buf+1, 
cannam@95: 					     buf, buf+1),
cannam@95: 			  NO_SLOW, 0, 0);
cannam@95:      if (!cldf) goto nada;
cannam@95: 
cannam@95:      X(ifree)(buf);
cannam@95: 
cannam@95:      pln = MKPLAN_DFT(P, &padt, apply);
cannam@95: 
cannam@95:      pln->n = n;
cannam@95:      pln->nb = nb;
cannam@95:      pln->w = 0;
cannam@95:      pln->W = 0;
cannam@95:      pln->cldf = cldf;
cannam@95:      pln->is = p->sz->dims[0].is;
cannam@95:      pln->os = p->sz->dims[0].os;
cannam@95: 
cannam@95:      X(ops_add)(&cldf->ops, &cldf->ops, &pln->super.super.ops);
cannam@95:      pln->super.super.ops.add += 4 * n + 2 * nb;
cannam@95:      pln->super.super.ops.mul += 8 * n + 4 * nb;
cannam@95:      pln->super.super.ops.other += 6 * (n + nb);
cannam@95: 
cannam@95:      return &(pln->super.super);
cannam@95: 
cannam@95:  nada:
cannam@95:      X(ifree0)(buf);
cannam@95:      X(plan_destroy_internal)(cldf);
cannam@95:      return (plan *)0;
cannam@95: }
cannam@95: 
cannam@95: 
cannam@95: static solver *mksolver(void)
cannam@95: {
cannam@95:      static const solver_adt sadt = { PROBLEM_DFT, mkplan, 0 };
cannam@95:      S *slv = MKSOLVER(S, &sadt);
cannam@95:      return &(slv->super);
cannam@95: }
cannam@95: 
cannam@95: void X(dft_bluestein_register)(planner *p)
cannam@95: {
cannam@95:      REGISTER_SOLVER(p, mksolver());
cannam@95: }