Chris@42: /*
Chris@42:  * Copyright (c) 2003, 2007-14 Matteo Frigo
Chris@42:  * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
Chris@42:  *
Chris@42:  * This program is free software; you can redistribute it and/or modify
Chris@42:  * it under the terms of the GNU General Public License as published by
Chris@42:  * the Free Software Foundation; either version 2 of the License, or
Chris@42:  * (at your option) any later version.
Chris@42:  *
Chris@42:  * This program is distributed in the hope that it will be useful,
Chris@42:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@42:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
Chris@42:  * GNU General Public License for more details.
Chris@42:  *
Chris@42:  * You should have received a copy of the GNU General Public License
Chris@42:  * along with this program; if not, write to the Free Software
Chris@42:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
Chris@42:  *
Chris@42:  */
Chris@42: 
Chris@42: #include "rdft.h"
Chris@42: 
Chris@42: typedef struct {
Chris@42:      solver super;
Chris@42:      rdft_kind kind;
Chris@42: } S;
Chris@42: 
Chris@42: typedef struct {
Chris@42:      plan_rdft super;
Chris@42:      twid *td;
Chris@42:      INT n, is, os;
Chris@42:      rdft_kind kind;
Chris@42: } P;
Chris@42: 
Chris@42: /***************************************************************************/
Chris@42: 
Chris@42: static void cdot_r2hc(INT n, const E *x, const R *w, R *or0, R *oi1)
Chris@42: {
Chris@42:      INT i;
Chris@42: 
Chris@42:      E rr = x[0], ri = 0;
Chris@42:      x += 1;
Chris@42:      for (i = 1; i + i < n; ++i) {
Chris@42: 	  rr += x[0] * w[0];
Chris@42: 	  ri += x[1] * w[1];
Chris@42: 	  x += 2; w += 2;
Chris@42:      }
Chris@42:      *or0 = rr;
Chris@42:      *oi1 = ri;
Chris@42: }
Chris@42: 
Chris@42: static void hartley_r2hc(INT n, const R *xr, INT xs, E *o, R *pr)
Chris@42: {
Chris@42:      INT i;
Chris@42:      E sr;
Chris@42:      o[0] = sr = xr[0]; o += 1;
Chris@42:      for (i = 1; i + i < n; ++i) {
Chris@42: 	  R a, b;
Chris@42: 	  a = xr[i * xs];
Chris@42: 	  b =  xr[(n - i) * xs];
Chris@42: 	  sr += (o[0] = a + b);
Chris@42: #if FFT_SIGN == -1
Chris@42: 	  o[1] = b - a;
Chris@42: #else
Chris@42: 	  o[1] = a - b;
Chris@42: #endif
Chris@42: 	  o += 2;
Chris@42:      }
Chris@42:      *pr = sr;
Chris@42: }
Chris@42: 		    
Chris@42: static void apply_r2hc(const plan *ego_, R *I, R *O)
Chris@42: {
Chris@42:      const P *ego = (const P *) ego_;
Chris@42:      INT i;
Chris@42:      INT n = ego->n, is = ego->is, os = ego->os;
Chris@42:      const R *W = ego->td->W;
Chris@42:      E *buf;
Chris@42:      size_t bufsz = n * sizeof(E);
Chris@42: 
Chris@42:      BUF_ALLOC(E *, buf, bufsz);
Chris@42:      hartley_r2hc(n, I, is, buf, O);
Chris@42: 
Chris@42:      for (i = 1; i + i < n; ++i) {
Chris@42: 	  cdot_r2hc(n, buf, W, O + i * os, O + (n - i) * os);
Chris@42: 	  W += n - 1;
Chris@42:      }
Chris@42: 
Chris@42:      BUF_FREE(buf, bufsz);
Chris@42: }
Chris@42: 
Chris@42: 
Chris@42: static void cdot_hc2r(INT n, const E *x, const R *w, R *or0, R *or1)
Chris@42: {
Chris@42:      INT i;
Chris@42: 
Chris@42:      E rr = x[0], ii = 0; 
Chris@42:      x += 1;
Chris@42:      for (i = 1; i + i < n; ++i) {
Chris@42: 	  rr += x[0] * w[0];
Chris@42: 	  ii += x[1] * w[1];
Chris@42: 	  x += 2; w += 2;
Chris@42:      }
Chris@42: #if FFT_SIGN == -1
Chris@42:      *or0 = rr - ii;
Chris@42:      *or1 = rr + ii;
Chris@42: #else
Chris@42:      *or0 = rr + ii;
Chris@42:      *or1 = rr - ii;
Chris@42: #endif
Chris@42: }
Chris@42: 
Chris@42: static void hartley_hc2r(INT n, const R *x, INT xs, E *o, R *pr)
Chris@42: {
Chris@42:      INT i;
Chris@42:      E sr;
Chris@42: 
Chris@42:      o[0] = sr = x[0]; o += 1;
Chris@42:      for (i = 1; i + i < n; ++i) {
Chris@42: 	  sr += (o[0] = x[i * xs] + x[i * xs]);
Chris@42: 	  o[1] = x[(n - i) * xs] + x[(n - i) * xs];
Chris@42: 	  o += 2;
Chris@42:      }
Chris@42:      *pr = sr;
Chris@42: }
Chris@42: 
Chris@42: static void apply_hc2r(const plan *ego_, R *I, R *O)		    
Chris@42: {
Chris@42:      const P *ego = (const P *) ego_;
Chris@42:      INT i;
Chris@42:      INT n = ego->n, is = ego->is, os = ego->os;
Chris@42:      const R *W = ego->td->W;
Chris@42:      E *buf;
Chris@42:      size_t bufsz = n * sizeof(E);
Chris@42: 
Chris@42:      BUF_ALLOC(E *, buf, bufsz);
Chris@42:      hartley_hc2r(n, I, is, buf, O);
Chris@42: 
Chris@42:      for (i = 1; i + i < n; ++i) {
Chris@42: 	  cdot_hc2r(n, buf, W, O + i * os, O + (n - i) * os);
Chris@42: 	  W += n - 1;
Chris@42:      }
Chris@42: 
Chris@42:      BUF_FREE(buf, bufsz);
Chris@42: }
Chris@42: 
Chris@42: 
Chris@42: /***************************************************************************/
Chris@42: 
Chris@42: static void awake(plan *ego_, enum wakefulness wakefulness)
Chris@42: {
Chris@42:      P *ego = (P *) ego_;
Chris@42:      static const tw_instr half_tw[] = {
Chris@42: 	  { TW_HALF, 1, 0 },
Chris@42: 	  { TW_NEXT, 1, 0 }
Chris@42:      };
Chris@42: 
Chris@42:      X(twiddle_awake)(wakefulness, &ego->td, half_tw, ego->n, ego->n,
Chris@42: 		      (ego->n - 1) / 2);
Chris@42: }
Chris@42: 
Chris@42: static void print(const plan *ego_, printer *p)
Chris@42: {
Chris@42:      const P *ego = (const P *) ego_;
Chris@42: 
Chris@42:      p->print(p, "(rdft-generic-%s-%D)", 
Chris@42: 	      ego->kind == R2HC ? "r2hc" : "hc2r", 
Chris@42: 	      ego->n);
Chris@42: }
Chris@42: 
Chris@42: static int applicable(const S *ego, const problem *p_, 
Chris@42: 		      const planner *plnr)
Chris@42: {
Chris@42:      const problem_rdft *p = (const problem_rdft *) p_;
Chris@42:      return (1
Chris@42: 	     && p->sz->rnk == 1
Chris@42: 	     && p->vecsz->rnk == 0
Chris@42: 	     && (p->sz->dims[0].n % 2) == 1 
Chris@42: 	     && CIMPLIES(NO_LARGE_GENERICP(plnr), p->sz->dims[0].n < GENERIC_MIN_BAD)
Chris@42: 	     && CIMPLIES(NO_SLOWP(plnr), p->sz->dims[0].n > GENERIC_MAX_SLOW)
Chris@42: 	     && X(is_prime)(p->sz->dims[0].n)
Chris@42: 	     && p->kind[0] == ego->kind
Chris@42: 	  );
Chris@42: }
Chris@42: 
Chris@42: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
Chris@42: {
Chris@42:      const S *ego = (const S *)ego_;
Chris@42:      const problem_rdft *p;
Chris@42:      P *pln;
Chris@42:      INT n;
Chris@42: 
Chris@42:      static const plan_adt padt = {
Chris@42: 	  X(rdft_solve), awake, print, X(plan_null_destroy)
Chris@42:      };
Chris@42: 
Chris@42:      if (!applicable(ego, p_, plnr))
Chris@42:           return (plan *)0;
Chris@42: 
Chris@42:      p = (const problem_rdft *) p_;
Chris@42:      pln = MKPLAN_RDFT(P, &padt, 
Chris@42: 		       R2HC_KINDP(p->kind[0]) ? apply_r2hc : apply_hc2r);
Chris@42: 
Chris@42:      pln->n = n = p->sz->dims[0].n;
Chris@42:      pln->is = p->sz->dims[0].is;
Chris@42:      pln->os = p->sz->dims[0].os;
Chris@42:      pln->td = 0;
Chris@42:      pln->kind = ego->kind;
Chris@42: 
Chris@42:      pln->super.super.ops.add = (n-1) * 2.5;
Chris@42:      pln->super.super.ops.mul = 0;
Chris@42:      pln->super.super.ops.fma = 0.5 * (n-1) * (n-1) ;
Chris@42: #if 0 /* these are nice pipelined sequential loads and should cost nothing */
Chris@42:      pln->super.super.ops.other = (n-1)*(2 + 1 + (n-1));  /* approximate */
Chris@42: #endif
Chris@42: 
Chris@42:      return &(pln->super.super);
Chris@42: }
Chris@42: 
Chris@42: static solver *mksolver(rdft_kind kind)
Chris@42: {
Chris@42:      static const solver_adt sadt = { PROBLEM_RDFT, mkplan, 0 };
Chris@42:      S *slv = MKSOLVER(S, &sadt);
Chris@42:      slv->kind = kind;
Chris@42:      return &(slv->super);
Chris@42: }
Chris@42: 
Chris@42: void X(rdft_generic_register)(planner *p)
Chris@42: {
Chris@42:      REGISTER_SOLVER(p, mksolver(R2HC));
Chris@42:      REGISTER_SOLVER(p, mksolver(HC2R));
Chris@42: }