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 "rdft.h"
cannam@95: 
cannam@95: typedef struct {
cannam@95:      solver super;
cannam@95:      rdft_kind kind;
cannam@95: } S;
cannam@95: 
cannam@95: typedef struct {
cannam@95:      plan_rdft super;
cannam@95:      twid *td;
cannam@95:      INT n, is, os;
cannam@95:      rdft_kind kind;
cannam@95: } P;
cannam@95: 
cannam@95: /***************************************************************************/
cannam@95: 
cannam@95: static void cdot_r2hc(INT n, const E *x, const R *w, R *or0, R *oi1)
cannam@95: {
cannam@95:      INT i;
cannam@95: 
cannam@95:      E rr = x[0], ri = 0;
cannam@95:      x += 1;
cannam@95:      for (i = 1; i + i < n; ++i) {
cannam@95: 	  rr += x[0] * w[0];
cannam@95: 	  ri += x[1] * w[1];
cannam@95: 	  x += 2; w += 2;
cannam@95:      }
cannam@95:      *or0 = rr;
cannam@95:      *oi1 = ri;
cannam@95: }
cannam@95: 
cannam@95: static void hartley_r2hc(INT n, const R *xr, INT xs, E *o, R *pr)
cannam@95: {
cannam@95:      INT i;
cannam@95:      E sr;
cannam@95:      o[0] = sr = xr[0]; o += 1;
cannam@95:      for (i = 1; i + i < n; ++i) {
cannam@95: 	  R a, b;
cannam@95: 	  a = xr[i * xs];
cannam@95: 	  b =  xr[(n - i) * xs];
cannam@95: 	  sr += (o[0] = a + b);
cannam@95: #if FFT_SIGN == -1
cannam@95: 	  o[1] = b - a;
cannam@95: #else
cannam@95: 	  o[1] = a - b;
cannam@95: #endif
cannam@95: 	  o += 2;
cannam@95:      }
cannam@95:      *pr = sr;
cannam@95: }
cannam@95: 		    
cannam@95: static void apply_r2hc(const plan *ego_, R *I, R *O)
cannam@95: {
cannam@95:      const P *ego = (const P *) ego_;
cannam@95:      INT i;
cannam@95:      INT n = ego->n, is = ego->is, os = ego->os;
cannam@95:      const R *W = ego->td->W;
cannam@95:      E *buf;
cannam@95:      size_t bufsz = n * sizeof(E);
cannam@95: 
cannam@95:      BUF_ALLOC(E *, buf, bufsz);
cannam@95:      hartley_r2hc(n, I, is, buf, O);
cannam@95: 
cannam@95:      for (i = 1; i + i < n; ++i) {
cannam@95: 	  cdot_r2hc(n, buf, W, O + i * os, O + (n - i) * os);
cannam@95: 	  W += n - 1;
cannam@95:      }
cannam@95: 
cannam@95:      BUF_FREE(buf, bufsz);
cannam@95: }
cannam@95: 
cannam@95: 
cannam@95: static void cdot_hc2r(INT n, const E *x, const R *w, R *or0, R *or1)
cannam@95: {
cannam@95:      INT i;
cannam@95: 
cannam@95:      E rr = x[0], ii = 0; 
cannam@95:      x += 1;
cannam@95:      for (i = 1; i + i < n; ++i) {
cannam@95: 	  rr += x[0] * w[0];
cannam@95: 	  ii += x[1] * w[1];
cannam@95: 	  x += 2; w += 2;
cannam@95:      }
cannam@95: #if FFT_SIGN == -1
cannam@95:      *or0 = rr - ii;
cannam@95:      *or1 = rr + ii;
cannam@95: #else
cannam@95:      *or0 = rr + ii;
cannam@95:      *or1 = rr - ii;
cannam@95: #endif
cannam@95: }
cannam@95: 
cannam@95: static void hartley_hc2r(INT n, const R *x, INT xs, E *o, R *pr)
cannam@95: {
cannam@95:      INT i;
cannam@95:      E sr;
cannam@95: 
cannam@95:      o[0] = sr = x[0]; o += 1;
cannam@95:      for (i = 1; i + i < n; ++i) {
cannam@95: 	  sr += (o[0] = x[i * xs] + x[i * xs]);
cannam@95: 	  o[1] = x[(n - i) * xs] + x[(n - i) * xs];
cannam@95: 	  o += 2;
cannam@95:      }
cannam@95:      *pr = sr;
cannam@95: }
cannam@95: 
cannam@95: static void apply_hc2r(const plan *ego_, R *I, R *O)		    
cannam@95: {
cannam@95:      const P *ego = (const P *) ego_;
cannam@95:      INT i;
cannam@95:      INT n = ego->n, is = ego->is, os = ego->os;
cannam@95:      const R *W = ego->td->W;
cannam@95:      E *buf;
cannam@95:      size_t bufsz = n * sizeof(E);
cannam@95: 
cannam@95:      BUF_ALLOC(E *, buf, bufsz);
cannam@95:      hartley_hc2r(n, I, is, buf, O);
cannam@95: 
cannam@95:      for (i = 1; i + i < n; ++i) {
cannam@95: 	  cdot_hc2r(n, buf, W, O + i * os, O + (n - i) * os);
cannam@95: 	  W += n - 1;
cannam@95:      }
cannam@95: 
cannam@95:      BUF_FREE(buf, bufsz);
cannam@95: }
cannam@95: 
cannam@95: 
cannam@95: /***************************************************************************/
cannam@95: 
cannam@95: static void awake(plan *ego_, enum wakefulness wakefulness)
cannam@95: {
cannam@95:      P *ego = (P *) ego_;
cannam@95:      static const tw_instr half_tw[] = {
cannam@95: 	  { TW_HALF, 1, 0 },
cannam@95: 	  { TW_NEXT, 1, 0 }
cannam@95:      };
cannam@95: 
cannam@95:      X(twiddle_awake)(wakefulness, &ego->td, half_tw, ego->n, ego->n,
cannam@95: 		      (ego->n - 1) / 2);
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: 
cannam@95:      p->print(p, "(rdft-generic-%s-%D)", 
cannam@95: 	      ego->kind == R2HC ? "r2hc" : "hc2r", 
cannam@95: 	      ego->n);
cannam@95: }
cannam@95: 
cannam@95: static int applicable(const S *ego, const problem *p_, 
cannam@95: 		      const planner *plnr)
cannam@95: {
cannam@95:      const problem_rdft *p = (const problem_rdft *) p_;
cannam@95:      return (1
cannam@95: 	     && p->sz->rnk == 1
cannam@95: 	     && p->vecsz->rnk == 0
cannam@95: 	     && (p->sz->dims[0].n % 2) == 1 
cannam@95: 	     && CIMPLIES(NO_LARGE_GENERICP(plnr), p->sz->dims[0].n < GENERIC_MIN_BAD)
cannam@95: 	     && CIMPLIES(NO_SLOWP(plnr), p->sz->dims[0].n > GENERIC_MAX_SLOW)
cannam@95: 	     && X(is_prime)(p->sz->dims[0].n)
cannam@95: 	     && p->kind[0] == ego->kind
cannam@95: 	  );
cannam@95: }
cannam@95: 
cannam@95: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
cannam@95: {
cannam@95:      const S *ego = (const S *)ego_;
cannam@95:      const problem_rdft *p;
cannam@95:      P *pln;
cannam@95:      INT n;
cannam@95: 
cannam@95:      static const plan_adt padt = {
cannam@95: 	  X(rdft_solve), awake, print, X(plan_null_destroy)
cannam@95:      };
cannam@95: 
cannam@95:      if (!applicable(ego, p_, plnr))
cannam@95:           return (plan *)0;
cannam@95: 
cannam@95:      p = (const problem_rdft *) p_;
cannam@95:      pln = MKPLAN_RDFT(P, &padt, 
cannam@95: 		       R2HC_KINDP(p->kind[0]) ? apply_r2hc : apply_hc2r);
cannam@95: 
cannam@95:      pln->n = n = p->sz->dims[0].n;
cannam@95:      pln->is = p->sz->dims[0].is;
cannam@95:      pln->os = p->sz->dims[0].os;
cannam@95:      pln->td = 0;
cannam@95:      pln->kind = ego->kind;
cannam@95: 
cannam@95:      pln->super.super.ops.add = (n-1) * 2.5;
cannam@95:      pln->super.super.ops.mul = 0;
cannam@95:      pln->super.super.ops.fma = 0.5 * (n-1) * (n-1) ;
cannam@95: #if 0 /* these are nice pipelined sequential loads and should cost nothing */
cannam@95:      pln->super.super.ops.other = (n-1)*(2 + 1 + (n-1));  /* approximate */
cannam@95: #endif
cannam@95: 
cannam@95:      return &(pln->super.super);
cannam@95: }
cannam@95: 
cannam@95: static solver *mksolver(rdft_kind kind)
cannam@95: {
cannam@95:      static const solver_adt sadt = { PROBLEM_RDFT, mkplan, 0 };
cannam@95:      S *slv = MKSOLVER(S, &sadt);
cannam@95:      slv->kind = kind;
cannam@95:      return &(slv->super);
cannam@95: }
cannam@95: 
cannam@95: void X(rdft_generic_register)(planner *p)
cannam@95: {
cannam@95:      REGISTER_SOLVER(p, mksolver(R2HC));
cannam@95:      REGISTER_SOLVER(p, mksolver(HC2R));
cannam@95: }