Chris@10: /*
Chris@10:  * Copyright (c) 2005 Matteo Frigo
Chris@10:  * Copyright (c) 2005 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: 
Chris@10: /* Do an R{E,O}DFT00 problem (of an odd length n) recursively via an
Chris@10:    R{E,O}DFT00 problem and an RDFT problem of half the length.
Chris@10: 
Chris@10:    This works by "logically" expanding the array to a real-even/odd DFT of
Chris@10:    length 2n-/+2 and then applying the split-radix algorithm.
Chris@10: 
Chris@10:    In this way, we can avoid having to pad to twice the length
Chris@10:    (ala redft00-r2hc-pad), saving a factor of ~2 for n=2^m+/-1,
Chris@10:    but don't incur the accuracy loss that the "ordinary" algorithm
Chris@10:    sacrifices (ala redft00-r2hc.c).
Chris@10: */
Chris@10: 
Chris@10: #include "reodft.h"
Chris@10: 
Chris@10: typedef struct {
Chris@10:      solver super;
Chris@10: } S;
Chris@10: 
Chris@10: typedef struct {
Chris@10:      plan_rdft super;
Chris@10:      plan *clde, *cldo;
Chris@10:      twid *td;
Chris@10:      INT is, os;
Chris@10:      INT n;
Chris@10:      INT vl;
Chris@10:      INT ivs, ovs;
Chris@10: } P;
Chris@10: 
Chris@10: /* redft00 */
Chris@10: static void apply_e(const plan *ego_, R *I, R *O)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      INT is = ego->is, os = ego->os;
Chris@10:      INT i, j, n = ego->n + 1, n2 = (n-1)/2;
Chris@10:      INT iv, vl = ego->vl;
Chris@10:      INT ivs = ego->ivs, ovs = ego->ovs;
Chris@10:      R *W = ego->td->W - 2;
Chris@10:      R *buf;
Chris@10: 
Chris@10:      buf = (R *) MALLOC(sizeof(R) * n2, BUFFERS);
Chris@10: 
Chris@10:      for (iv = 0; iv < vl; ++iv, I += ivs, O += ovs) {
Chris@10: 	  /* do size (n-1)/2 r2hc transform of odd-indexed elements
Chris@10: 	     with stride 4, "wrapping around" end of array with even
Chris@10: 	     boundary conditions */
Chris@10: 	  for (j = 0, i = 1; i < n; i += 4)
Chris@10: 	       buf[j++] = I[is * i];
Chris@10: 	  for (i = 2*n-2-i; i > 0; i -= 4)
Chris@10: 	       buf[j++] = I[is * i];
Chris@10: 	  {
Chris@10: 	       plan_rdft *cld = (plan_rdft *) ego->cldo;
Chris@10: 	       cld->apply((plan *) cld, buf, buf);
Chris@10: 	  }
Chris@10: 
Chris@10: 	  /* do size (n+1)/2 redft00 of the even-indexed elements,
Chris@10: 	     writing to O: */
Chris@10: 	  {
Chris@10: 	       plan_rdft *cld = (plan_rdft *) ego->clde;
Chris@10: 	       cld->apply((plan *) cld, I, O);
Chris@10: 	  }
Chris@10: 
Chris@10: 	  /* combine the results with the twiddle factors to get output */
Chris@10: 	  { /* DC element */
Chris@10: 	       E b20 = O[0], b0 = K(2.0) * buf[0];
Chris@10: 	       O[0] = b20 + b0;
Chris@10: 	       O[2*(n2*os)] = b20 - b0;
Chris@10: 	       /* O[n2*os] = O[n2*os]; */
Chris@10: 	  }
Chris@10: 	  for (i = 1; i < n2 - i; ++i) {
Chris@10: 	       E ap, am, br, bi, wr, wi, wbr, wbi;
Chris@10: 	       br = buf[i];
Chris@10: 	       bi = buf[n2 - i];
Chris@10: 	       wr = W[2*i];
Chris@10: 	       wi = W[2*i+1];
Chris@10: #if FFT_SIGN == -1
Chris@10: 	       wbr = K(2.0) * (wr*br + wi*bi);
Chris@10: 	       wbi = K(2.0) * (wr*bi - wi*br);
Chris@10: #else
Chris@10: 	       wbr = K(2.0) * (wr*br - wi*bi);
Chris@10: 	       wbi = K(2.0) * (wr*bi + wi*br);
Chris@10: #endif
Chris@10: 	       ap = O[i*os];
Chris@10: 	       O[i*os] = ap + wbr;
Chris@10: 	       O[(2*n2 - i)*os] = ap - wbr;
Chris@10: 	       am = O[(n2 - i)*os];
Chris@10: #if FFT_SIGN == -1
Chris@10: 	       O[(n2 - i)*os] = am - wbi;
Chris@10: 	       O[(n2 + i)*os] = am + wbi;
Chris@10: #else
Chris@10: 	       O[(n2 - i)*os] = am + wbi;
Chris@10: 	       O[(n2 + i)*os] = am - wbi;
Chris@10: #endif
Chris@10: 	  }
Chris@10: 	  if (i == n2 - i) { /* Nyquist element */
Chris@10: 	       E ap, wbr;
Chris@10: 	       wbr = K(2.0) * (W[2*i] * buf[i]);
Chris@10: 	       ap = O[i*os];
Chris@10: 	       O[i*os] = ap + wbr;
Chris@10: 	       O[(2*n2 - i)*os] = ap - wbr;
Chris@10: 	  }
Chris@10:      }
Chris@10: 
Chris@10:      X(ifree)(buf);
Chris@10: }
Chris@10: 
Chris@10: /* rodft00 */
Chris@10: static void apply_o(const plan *ego_, R *I, R *O)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      INT is = ego->is, os = ego->os;
Chris@10:      INT i, j, n = ego->n - 1, n2 = (n+1)/2;
Chris@10:      INT iv, vl = ego->vl;
Chris@10:      INT ivs = ego->ivs, ovs = ego->ovs;
Chris@10:      R *W = ego->td->W - 2;
Chris@10:      R *buf;
Chris@10: 
Chris@10:      buf = (R *) MALLOC(sizeof(R) * n2, BUFFERS);
Chris@10: 
Chris@10:      for (iv = 0; iv < vl; ++iv, I += ivs, O += ovs) {
Chris@10: 	  /* do size (n+1)/2 r2hc transform of even-indexed elements
Chris@10: 	     with stride 4, "wrapping around" end of array with odd
Chris@10: 	     boundary conditions */
Chris@10: 	  for (j = 0, i = 0; i < n; i += 4)
Chris@10: 	       buf[j++] = I[is * i];
Chris@10: 	  for (i = 2*n-i; i > 0; i -= 4)
Chris@10: 	       buf[j++] = -I[is * i];
Chris@10: 	  {
Chris@10: 	       plan_rdft *cld = (plan_rdft *) ego->cldo;
Chris@10: 	       cld->apply((plan *) cld, buf, buf);
Chris@10: 	  }
Chris@10: 
Chris@10: 	  /* do size (n-1)/2 rodft00 of the odd-indexed elements,
Chris@10: 	     writing to O: */
Chris@10: 	  {
Chris@10: 	       plan_rdft *cld = (plan_rdft *) ego->clde;
Chris@10: 	       if (I == O) {
Chris@10: 		    /* can't use I+is and I, subplan would lose in-placeness */
Chris@10: 		    cld->apply((plan *) cld, I + is, I + is);
Chris@10: 		    /* we could maybe avoid this copy by modifying the
Chris@10: 		       twiddle loop, but currently I can't be bothered. */
Chris@10: 		    A(is >= os);
Chris@10: 		    for (i = 0; i < n2-1; ++i)
Chris@10: 			 O[os*i] = I[is*(i+1)];
Chris@10: 	       }
Chris@10: 	       else
Chris@10: 		    cld->apply((plan *) cld, I + is, O);
Chris@10: 	  }
Chris@10: 
Chris@10: 	  /* combine the results with the twiddle factors to get output */
Chris@10: 	  O[(n2-1)*os] = K(2.0) * buf[0];
Chris@10: 	  for (i = 1; i < n2 - i; ++i) {
Chris@10: 	       E ap, am, br, bi, wr, wi, wbr, wbi;
Chris@10: 	       br = buf[i];
Chris@10: 	       bi = buf[n2 - i];
Chris@10: 	       wr = W[2*i];
Chris@10: 	       wi = W[2*i+1];
Chris@10: #if FFT_SIGN == -1
Chris@10: 	       wbr = K(2.0) * (wr*br + wi*bi);
Chris@10: 	       wbi = K(2.0) * (wi*br - wr*bi);
Chris@10: #else
Chris@10: 	       wbr = K(2.0) * (wr*br - wi*bi);
Chris@10: 	       wbi = K(2.0) * (wr*bi + wi*br);
Chris@10: #endif
Chris@10: 	       ap = O[(i-1)*os];
Chris@10: 	       O[(i-1)*os] = wbi + ap;
Chris@10: 	       O[(2*n2-1 - i)*os] = wbi - ap;
Chris@10: 	       am = O[(n2-1 - i)*os];
Chris@10: #if FFT_SIGN == -1
Chris@10: 	       O[(n2-1 - i)*os] = wbr + am;
Chris@10: 	       O[(n2-1 + i)*os] = wbr - am;
Chris@10: #else
Chris@10: 	       O[(n2-1 - i)*os] = wbr + am;
Chris@10: 	       O[(n2-1 + i)*os] = wbr - am;
Chris@10: #endif
Chris@10: 	  }
Chris@10: 	  if (i == n2 - i) { /* Nyquist element */
Chris@10: 	       E ap, wbi;
Chris@10: 	       wbi = K(2.0) * (W[2*i+1] * buf[i]);
Chris@10: 	       ap = O[(i-1)*os];
Chris@10: 	       O[(i-1)*os] = wbi + ap;
Chris@10: 	       O[(2*n2-1 - i)*os] = wbi - ap;
Chris@10: 	  }
Chris@10:      }
Chris@10: 
Chris@10:      X(ifree)(buf);
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 reodft00e_tw[] = {
Chris@10:           { TW_COS, 1, 1 },
Chris@10:           { TW_SIN, 1, 1 },
Chris@10:           { TW_NEXT, 1, 0 }
Chris@10:      };
Chris@10: 
Chris@10:      X(plan_awake)(ego->clde, wakefulness);
Chris@10:      X(plan_awake)(ego->cldo, wakefulness);
Chris@10:      X(twiddle_awake)(wakefulness, &ego->td, reodft00e_tw, 
Chris@10: 		      2*ego->n, 1, ego->n/4);
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->cldo);
Chris@10:      X(plan_destroy_internal)(ego->clde);
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:      if (ego->super.apply == apply_e)
Chris@10: 	  p->print(p, "(redft00e-splitradix-%D%v%(%p%)%(%p%))", 
Chris@10: 		   ego->n + 1, ego->vl, ego->clde, ego->cldo);
Chris@10:      else
Chris@10: 	  p->print(p, "(rodft00e-splitradix-%D%v%(%p%)%(%p%))", 
Chris@10: 		   ego->n - 1, ego->vl, ego->clde, ego->cldo);
Chris@10: }
Chris@10: 
Chris@10: static int applicable0(const solver *ego_, const problem *p_)
Chris@10: {
Chris@10:      const problem_rdft *p = (const problem_rdft *) p_;
Chris@10:      UNUSED(ego_);
Chris@10: 
Chris@10:      return (1
Chris@10: 	     && p->sz->rnk == 1
Chris@10: 	     && p->vecsz->rnk <= 1
Chris@10: 	     && (p->kind[0] == REDFT00 || p->kind[0] == RODFT00)
Chris@10: 	     && p->sz->dims[0].n > 1  /* don't create size-0 sub-plans */
Chris@10: 	     && p->sz->dims[0].n % 2  /* odd: 4 divides "logical" DFT */
Chris@10: 	     && (p->I != p->O || p->vecsz->rnk == 0
Chris@10: 		 || p->vecsz->dims[0].is == p->vecsz->dims[0].os)
Chris@10: 	     && (p->kind[0] != RODFT00 || p->I != p->O || 
Chris@10: 		 p->sz->dims[0].is >= p->sz->dims[0].os) /* laziness */
Chris@10: 	  );
Chris@10: }
Chris@10: 
Chris@10: static int applicable(const solver *ego, const problem *p, const planner *plnr)
Chris@10: {
Chris@10:      return (!NO_SLOWP(plnr) && applicable0(ego, p));
Chris@10: }
Chris@10: 
Chris@10: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
Chris@10: {
Chris@10:      P *pln;
Chris@10:      const problem_rdft *p;
Chris@10:      plan *clde, *cldo;
Chris@10:      R *buf;
Chris@10:      INT n, n0;
Chris@10:      opcnt ops;
Chris@10:      int inplace_odd;
Chris@10: 
Chris@10:      static const plan_adt padt = {
Chris@10: 	  X(rdft_solve), awake, print, destroy
Chris@10:      };
Chris@10: 
Chris@10:      if (!applicable(ego_, p_, plnr))
Chris@10:           return (plan *)0;
Chris@10: 
Chris@10:      p = (const problem_rdft *) p_;
Chris@10: 
Chris@10:      n = (n0 = p->sz->dims[0].n) + (p->kind[0] == REDFT00 ? (INT)-1 : (INT)1);
Chris@10:      A(n > 0 && n % 2 == 0);
Chris@10:      buf = (R *) MALLOC(sizeof(R) * (n/2), BUFFERS);
Chris@10: 
Chris@10:      inplace_odd = p->kind[0]==RODFT00 && p->I == p->O;
Chris@10:      clde = X(mkplan_d)(plnr, X(mkproblem_rdft_1_d)(
Chris@10: 			     X(mktensor_1d)(n0-n/2, 2*p->sz->dims[0].is, 
Chris@10: 					    inplace_odd ? p->sz->dims[0].is
Chris@10: 					    : p->sz->dims[0].os), 
Chris@10: 			     X(mktensor_0d)(), 
Chris@10: 			     TAINT(p->I 
Chris@10: 				   + p->sz->dims[0].is * (p->kind[0]==RODFT00),
Chris@10: 				   p->vecsz->rnk ? p->vecsz->dims[0].is : 0),
Chris@10: 			     TAINT(p->O
Chris@10: 				   + p->sz->dims[0].is * inplace_odd,
Chris@10: 				   p->vecsz->rnk ? p->vecsz->dims[0].os : 0),
Chris@10: 			     p->kind[0]));
Chris@10:      if (!clde) {
Chris@10: 	  X(ifree)(buf);
Chris@10:           return (plan *)0;
Chris@10:      }
Chris@10: 
Chris@10:      cldo = X(mkplan_d)(plnr, X(mkproblem_rdft_1_d)(
Chris@10: 			     X(mktensor_1d)(n/2, 1, 1), 
Chris@10: 			     X(mktensor_0d)(), 
Chris@10: 			     buf, buf, R2HC));
Chris@10:      X(ifree)(buf);
Chris@10:      if (!cldo)
Chris@10:           return (plan *)0;
Chris@10: 
Chris@10:      pln = MKPLAN_RDFT(P, &padt, p->kind[0] == REDFT00 ? apply_e : apply_o);
Chris@10: 
Chris@10:      pln->n = n;
Chris@10:      pln->is = p->sz->dims[0].is;
Chris@10:      pln->os = p->sz->dims[0].os;
Chris@10:      pln->clde = clde;
Chris@10:      pln->cldo = cldo;
Chris@10:      pln->td = 0;
Chris@10: 
Chris@10:      X(tensor_tornk1)(p->vecsz, &pln->vl, &pln->ivs, &pln->ovs);
Chris@10:      
Chris@10:      X(ops_zero)(&ops);
Chris@10:      ops.other = n/2;
Chris@10:      ops.add = (p->kind[0]==REDFT00 ? (INT)2 : (INT)0) +
Chris@10: 	  (n/2-1)/2 * 6 + ((n/2)%2==0) * 2;
Chris@10:      ops.mul = 1 + (n/2-1)/2 * 6 + ((n/2)%2==0) * 2;
Chris@10: 
Chris@10:      /* tweak ops.other so that r2hc-pad is used for small sizes, which
Chris@10: 	seems to be a lot faster on my machine: */
Chris@10:      ops.other += 256;
Chris@10: 
Chris@10:      X(ops_zero)(&pln->super.super.ops);
Chris@10:      X(ops_madd2)(pln->vl, &ops, &pln->super.super.ops);
Chris@10:      X(ops_madd2)(pln->vl, &clde->ops, &pln->super.super.ops);
Chris@10:      X(ops_madd2)(pln->vl, &cldo->ops, &pln->super.super.ops);
Chris@10: 
Chris@10:      return &(pln->super.super);
Chris@10: }
Chris@10: 
Chris@10: /* constructor */
Chris@10: static solver *mksolver(void)
Chris@10: {
Chris@10:      static const solver_adt sadt = { PROBLEM_RDFT, mkplan, 0 };
Chris@10:      S *slv = MKSOLVER(S, &sadt);
Chris@10:      return &(slv->super);
Chris@10: }
Chris@10: 
Chris@10: void X(reodft00e_splitradix_register)(planner *p)
Chris@10: {
Chris@10:      REGISTER_SOLVER(p, mksolver());
Chris@10: }