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:      twid *td;
Chris@10:      INT n, is, os;
Chris@10: } P;
Chris@10: 
Chris@10: 
Chris@10: static void cdot(INT n, const E *x, const R *w, 
Chris@10: 		 R *or0, R *oi0, R *or1, R *oi1)
Chris@10: {
Chris@10:      INT i;
Chris@10: 
Chris@10:      E rr = x[0], ri = 0, ir = x[1], ii = 0; 
Chris@10:      x += 2;
Chris@10:      for (i = 1; i + i < n; ++i) {
Chris@10: 	  rr += x[0] * w[0];
Chris@10: 	  ir += x[1] * w[0];
Chris@10: 	  ri += x[2] * w[1];
Chris@10: 	  ii += x[3] * w[1];
Chris@10: 	  x += 4; w += 2;
Chris@10:      }
Chris@10:      *or0 = rr + ii;
Chris@10:      *oi0 = ir - ri;
Chris@10:      *or1 = rr - ii;
Chris@10:      *oi1 = ir + ri;
Chris@10: }
Chris@10: 
Chris@10: static void hartley(INT n, const R *xr, const R *xi, INT xs, E *o,
Chris@10: 		    R *pr, R *pi)
Chris@10: {
Chris@10:      INT i;
Chris@10:      E sr, si;
Chris@10:      o[0] = sr = xr[0]; o[1] = si = xi[0]; o += 2;
Chris@10:      for (i = 1; i + i < n; ++i) {
Chris@10: 	  sr += (o[0] = xr[i * xs] + xr[(n - i) * xs]);
Chris@10: 	  si += (o[1] = xi[i * xs] + xi[(n - i) * xs]);
Chris@10: 	  o[2] = xr[i * xs] - xr[(n - i) * xs];
Chris@10: 	  o[3] = xi[i * xs] - xi[(n - i) * xs];
Chris@10: 	  o += 4;
Chris@10:      }
Chris@10:      *pr = sr;
Chris@10:      *pi = si;
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;
Chris@10:      INT n = ego->n, is = ego->is, os = ego->os;
Chris@10:      const R *W = ego->td->W;
Chris@10:      E *buf;
Chris@10:      size_t bufsz = n * 2 * sizeof(E);
Chris@10: 
Chris@10:      BUF_ALLOC(E *, buf, bufsz);
Chris@10:      hartley(n, ri, ii, is, buf, ro, io);
Chris@10: 
Chris@10:      for (i = 1; i + i < n; ++i) {
Chris@10: 	  cdot(n, buf, W,
Chris@10: 	       ro + i * os, io + i * os,
Chris@10: 	       ro + (n - i) * os, io + (n - i) * os);
Chris@10: 	  W += n - 1;
Chris@10:      }
Chris@10: 
Chris@10:      BUF_FREE(buf, bufsz);
Chris@10: }
Chris@10: 
Chris@10: static void awake(plan *ego_, enum wakefulness wakefulness)
Chris@10: {
Chris@10:      P *ego = (P *) ego_;
Chris@10:      static const tw_instr half_tw[] = {
Chris@10: 	  { TW_HALF, 1, 0 },
Chris@10: 	  { TW_NEXT, 1, 0 }
Chris@10:      };
Chris@10: 
Chris@10:      X(twiddle_awake)(wakefulness, &ego->td, half_tw, ego->n, ego->n,
Chris@10: 		      (ego->n - 1) / 2);
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: 
Chris@10:      p->print(p, "(dft-generic-%D)", ego->n);
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: 
Chris@10:      return (1
Chris@10: 	     && p->sz->rnk == 1
Chris@10: 	     && p->vecsz->rnk == 0
Chris@10: 	     && (p->sz->dims[0].n % 2) == 1 
Chris@10: 	     && CIMPLIES(NO_LARGE_GENERICP(plnr), p->sz->dims[0].n < GENERIC_MIN_BAD)
Chris@10: 	     && CIMPLIES(NO_SLOWP(plnr), p->sz->dims[0].n > GENERIC_MAX_SLOW)
Chris@10: 	     && X(is_prime)(p->sz->dims[0].n)
Chris@10: 	  );
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;
Chris@10:      P *pln;
Chris@10:      INT n;
Chris@10: 
Chris@10:      static const plan_adt padt = {
Chris@10: 	  X(dft_solve), awake, print, X(plan_null_destroy)
Chris@10:      };
Chris@10: 
Chris@10:      if (!applicable(ego, p_, plnr))
Chris@10:           return (plan *)0;
Chris@10: 
Chris@10:      pln = MKPLAN_DFT(P, &padt, apply);
Chris@10: 
Chris@10:      p = (const problem_dft *) p_;
Chris@10:      pln->n = n = p->sz->dims[0].n;
Chris@10:      pln->is = p->sz->dims[0].is;
Chris@10:      pln->os = p->sz->dims[0].os;
Chris@10:      pln->td = 0;
Chris@10: 
Chris@10:      pln->super.super.ops.add = (n-1) * 5;
Chris@10:      pln->super.super.ops.mul = 0;
Chris@10:      pln->super.super.ops.fma = (n-1) * (n-1) ;
Chris@10: #if 0 /* these are nice pipelined sequential loads and should cost nothing */
Chris@10:      pln->super.super.ops.other = (n-1)*(4 + 1 + 2 * (n-1));  /* approximate */
Chris@10: #endif
Chris@10: 
Chris@10:      return &(pln->super.super);
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_generic_register)(planner *p)
Chris@10: {
Chris@10:      REGISTER_SOLVER(p, mksolver());
Chris@10: }