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: /* express a hc2hc problem in terms of rdft + multiplication by
Chris@10:    twiddle factors */
Chris@10: 
Chris@10: #include "hc2hc.h"
Chris@10: 
Chris@10: typedef hc2hc_solver S;
Chris@10: 
Chris@10: typedef struct {
Chris@10:      plan_hc2hc super;
Chris@10: 
Chris@10:      INT r, m, s, vl, vs, mstart1, mcount1;
Chris@10:      plan *cld0;
Chris@10:      plan *cld;
Chris@10:      twid *td;
Chris@10: } P;
Chris@10: 
Chris@10: 
Chris@10: /**************************************************************/
Chris@10: static void mktwiddle(P *ego, enum wakefulness wakefulness)
Chris@10: {
Chris@10:      static const tw_instr tw[] = { { TW_HALF, 0, 0 }, { TW_NEXT, 1, 0 } };
Chris@10: 
Chris@10:      /* note that R and M are swapped, to allow for sequential
Chris@10: 	access both to data and twiddles */
Chris@10:      X(twiddle_awake)(wakefulness, &ego->td, tw, 
Chris@10: 		      ego->r * ego->m, ego->m, ego->r);
Chris@10: }
Chris@10: 
Chris@10: static void bytwiddle(const P *ego, R *IO, R sign)
Chris@10: {
Chris@10:      INT i, j, k;
Chris@10:      INT r = ego->r, m = ego->m, s = ego->s, vl = ego->vl, vs = ego->vs;
Chris@10:      INT ms = m * s;
Chris@10:      INT mstart1 = ego->mstart1, mcount1 = ego->mcount1;
Chris@10:      INT wrem = 2 * ((m-1)/2 - mcount1);
Chris@10: 
Chris@10:      for (i = 0; i < vl; ++i, IO += vs) {
Chris@10: 	  const R *W = ego->td->W;
Chris@10: 
Chris@10: 	  A(m % 2 == 1);
Chris@10: 	  for (k = 1, W += (m - 1) + 2*(mstart1-1); k < r; ++k) {
Chris@10: 	       /* pr := IO + (j + mstart1) * s + k * ms */
Chris@10: 	       R *pr = IO + mstart1 * s + k * ms;
Chris@10: 
Chris@10: 	       /* pi := IO + (m - j - mstart1) * s + k * ms */
Chris@10: 	       R *pi = IO - mstart1 * s + (k + 1) * ms;
Chris@10: 
Chris@10: 	       for (j = 0; j < mcount1; ++j, pr += s, pi -= s) {
Chris@10: 		    E xr = *pr;
Chris@10: 		    E xi = *pi;
Chris@10: 		    E wr = W[0];
Chris@10: 		    E wi = sign * W[1];
Chris@10: 		    *pr = xr * wr - xi * wi;
Chris@10: 		    *pi = xi * wr + xr * wi;
Chris@10: 		    W += 2;
Chris@10: 	       }
Chris@10: 	       W += wrem;
Chris@10: 	  }
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: static void swapri(R *IO, INT r, INT m, INT s, INT jstart, INT jend)
Chris@10: {
Chris@10:      INT k;
Chris@10:      INT ms = m * s;
Chris@10:      INT js = jstart * s;
Chris@10:      for (k = 0; k + k < r; ++k) {
Chris@10: 	  /* pr := IO + (m - j) * s + k * ms */
Chris@10: 	  R *pr = IO + (k + 1) * ms - js;
Chris@10: 	  /* pi := IO + (m - j) * s + (r - 1 - k) * ms */
Chris@10: 	  R *pi = IO + (r - k) * ms - js;
Chris@10: 	  INT j;
Chris@10: 	  for (j = jstart; j < jend; j += 1, pr -= s, pi -= s) {
Chris@10: 	       R t = *pr;
Chris@10: 	       *pr = *pi;
Chris@10: 	       *pi = t;
Chris@10: 	  }
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: static void reorder_dit(const P *ego, R *IO)
Chris@10: {
Chris@10:      INT i, k;
Chris@10:      INT r = ego->r, m = ego->m, s = ego->s, vl = ego->vl, vs = ego->vs;
Chris@10:      INT ms = m * s;
Chris@10:      INT mstart1 = ego->mstart1, mend1 = mstart1 + ego->mcount1;
Chris@10: 
Chris@10:      for (i = 0; i < vl; ++i, IO += vs) {
Chris@10: 	  for (k = 1; k + k < r; ++k) {
Chris@10: 	       R *p0 = IO + k * ms;
Chris@10: 	       R *p1 = IO + (r - k) * ms;
Chris@10: 	       INT j;
Chris@10: 
Chris@10: 	       for (j = mstart1; j < mend1; ++j) {
Chris@10: 		    E rp, ip, im, rm;
Chris@10: 		    rp = p0[j * s];
Chris@10: 		    im = p1[ms - j * s];
Chris@10: 		    rm = p1[j * s];
Chris@10: 		    ip = p0[ms - j * s];
Chris@10: 		    p0[j * s] = rp - im;
Chris@10: 		    p1[ms - j * s] = rp + im;
Chris@10: 		    p1[j * s] = rm - ip;
Chris@10: 		    p0[ms - j * s] = ip + rm;
Chris@10: 	       }
Chris@10: 	  }
Chris@10: 
Chris@10: 	  swapri(IO, r, m, s, mstart1, mend1);
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: static void reorder_dif(const P *ego, R *IO)
Chris@10: {
Chris@10:      INT i, k;
Chris@10:      INT r = ego->r, m = ego->m, s = ego->s, vl = ego->vl, vs = ego->vs;
Chris@10:      INT ms = m * s;
Chris@10:      INT mstart1 = ego->mstart1, mend1 = mstart1 + ego->mcount1;
Chris@10: 
Chris@10:      for (i = 0; i < vl; ++i, IO += vs) {
Chris@10: 	  swapri(IO, r, m, s, mstart1, mend1);
Chris@10: 
Chris@10: 	  for (k = 1; k + k < r; ++k) {
Chris@10: 	       R *p0 = IO + k * ms;
Chris@10: 	       R *p1 = IO + (r - k) * ms;
Chris@10: 	       const R half = K(0.5);
Chris@10: 	       INT j;
Chris@10: 
Chris@10: 	       for (j = mstart1; j < mend1; ++j) {
Chris@10: 		    E rp, ip, im, rm;
Chris@10: 		    rp = half * p0[j * s];
Chris@10: 		    im = half * p1[ms - j * s];
Chris@10: 		    rm = half * p1[j * s];
Chris@10: 		    ip = half * p0[ms - j * s];
Chris@10: 		    p0[j * s] = rp + im;
Chris@10: 		    p1[ms - j * s] = im - rp;
Chris@10: 		    p1[j * s] = rm + ip;
Chris@10: 		    p0[ms - j * s] = ip - rm;
Chris@10: 	       }
Chris@10: 	  }
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: static int applicable(rdft_kind kind, INT r, INT m, const planner *plnr)
Chris@10: {
Chris@10:      return (1 
Chris@10: 	     && (kind == R2HC || kind == HC2R)
Chris@10: 	     && (m % 2)
Chris@10: 	     && (r % 2)
Chris@10: 	     && !NO_SLOWP(plnr)
Chris@10: 	  );
Chris@10: }
Chris@10: 
Chris@10: /**************************************************************/
Chris@10: 
Chris@10: static void apply_dit(const plan *ego_, R *IO)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      INT start;
Chris@10:      plan_rdft *cld, *cld0;
Chris@10: 
Chris@10:      bytwiddle(ego, IO, K(-1.0));
Chris@10: 
Chris@10:      cld0 = (plan_rdft *) ego->cld0;
Chris@10:      cld0->apply(ego->cld0, IO, IO);
Chris@10: 
Chris@10:      start = ego->mstart1 * ego->s;
Chris@10:      cld = (plan_rdft *) ego->cld;
Chris@10:      cld->apply(ego->cld, IO + start, IO + start);
Chris@10: 
Chris@10:      reorder_dit(ego, IO);
Chris@10: }
Chris@10: 
Chris@10: static void apply_dif(const plan *ego_, R *IO)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      INT start;
Chris@10:      plan_rdft *cld, *cld0;
Chris@10: 
Chris@10:      reorder_dif(ego, IO);
Chris@10: 
Chris@10:      cld0 = (plan_rdft *) ego->cld0;
Chris@10:      cld0->apply(ego->cld0, IO, IO);
Chris@10: 
Chris@10:      start = ego->mstart1 * ego->s;
Chris@10:      cld = (plan_rdft *) ego->cld;
Chris@10:      cld->apply(ego->cld, IO + start, IO + start);
Chris@10: 
Chris@10:      bytwiddle(ego, IO, K(1.0));
Chris@10: }
Chris@10: 
Chris@10: 
Chris@10: static void awake(plan *ego_, enum wakefulness wakefulness)
Chris@10: {
Chris@10:      P *ego = (P *) ego_;
Chris@10:      X(plan_awake)(ego->cld0, wakefulness);
Chris@10:      X(plan_awake)(ego->cld, wakefulness);
Chris@10:      mktwiddle(ego, wakefulness);
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->cld);
Chris@10:      X(plan_destroy_internal)(ego->cld0);
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:      p->print(p, "(hc2hc-generic-%s-%D-%D%v%(%p%)%(%p%))", 
Chris@10: 	      ego->super.apply == apply_dit ? "dit" : "dif",
Chris@10: 	      ego->r, ego->m, ego->vl, ego->cld0, ego->cld);
Chris@10: }
Chris@10: 
Chris@10: static plan *mkcldw(const hc2hc_solver *ego_, 
Chris@10: 		    rdft_kind kind, INT r, INT m, INT s, INT vl, INT vs, 
Chris@10: 		    INT mstart, INT mcount,
Chris@10: 		    R *IO, planner *plnr)
Chris@10: {
Chris@10:      P *pln;
Chris@10:      plan *cld0 = 0, *cld = 0;
Chris@10:      INT mstart1, mcount1, mstride;
Chris@10: 
Chris@10:      static const plan_adt padt = {
Chris@10: 	  0, awake, print, destroy
Chris@10:      };
Chris@10: 
Chris@10:      UNUSED(ego_);
Chris@10: 
Chris@10:      A(mstart >= 0 && mcount > 0 && mstart + mcount <= (m+2)/2);
Chris@10: 
Chris@10:      if (!applicable(kind, r, m, plnr))
Chris@10:           return (plan *)0;
Chris@10: 
Chris@10:      A(m % 2);
Chris@10:      mstart1 = mstart + (mstart == 0);
Chris@10:      mcount1 = mcount - (mstart == 0);
Chris@10:      mstride = m - (mstart + mcount - 1) - mstart1;
Chris@10: 
Chris@10:      /* 0th (DC) transform (vl of these), if mstart == 0 */
Chris@10:      cld0 = X(mkplan_d)(plnr, 
Chris@10: 			X(mkproblem_rdft_1_d)(
Chris@10: 			     mstart == 0 ? X(mktensor_1d)(r, m * s, m * s)
Chris@10: 			     : X(mktensor_0d)(),
Chris@10: 			     X(mktensor_1d)(vl, vs, vs),
Chris@10: 			     IO, IO, kind)
Chris@10: 			);
Chris@10:      if (!cld0) goto nada;
Chris@10: 
Chris@10:      /* twiddle transforms: there are 2 x mcount1 x vl of these
Chris@10: 	(where 2 corresponds to the real and imaginary parts) ...
Chris@10:         the 2 x mcount1 loops are combined if mstart=0 and mcount=(m+2)/2. */
Chris@10:      cld = X(mkplan_d)(plnr, 
Chris@10: 			X(mkproblem_rdft_1_d)(
Chris@10: 			     X(mktensor_1d)(r, m * s, m * s),
Chris@10: 			     X(mktensor_3d)(2, mstride * s, mstride * s,
Chris@10: 					    mcount1, s, s, 
Chris@10: 					    vl, vs, vs),
Chris@10: 			     IO + s * mstart1, IO + s * mstart1, kind)
Chris@10: 	                );
Chris@10:      if (!cld) goto nada;
Chris@10:      
Chris@10:      pln = MKPLAN_HC2HC(P, &padt, (kind == R2HC) ? apply_dit : apply_dif);
Chris@10:      pln->cld = cld;
Chris@10:      pln->cld0 = cld0;
Chris@10:      pln->r = r;
Chris@10:      pln->m = m;
Chris@10:      pln->s = s;
Chris@10:      pln->vl = vl;
Chris@10:      pln->vs = vs;
Chris@10:      pln->td = 0;
Chris@10:      pln->mstart1 = mstart1;
Chris@10:      pln->mcount1 = mcount1;
Chris@10: 
Chris@10:      {
Chris@10: 	  double n0 = 0.5 * (r - 1) * (2 * mcount1) * vl;
Chris@10: 	  pln->super.super.ops = cld->ops;
Chris@10: 	  pln->super.super.ops.mul += (kind == R2HC ? 5.0 : 7.0) * n0;
Chris@10: 	  pln->super.super.ops.add += 4.0 * n0;
Chris@10: 	  pln->super.super.ops.other += 11.0 * n0;
Chris@10:      }
Chris@10:      return &(pln->super.super);
Chris@10: 
Chris@10:  nada:
Chris@10:      X(plan_destroy_internal)(cld);
Chris@10:      X(plan_destroy_internal)(cld0);
Chris@10:      return (plan *) 0;
Chris@10: }
Chris@10: 
Chris@10: static void regsolver(planner *plnr, INT r)
Chris@10: {
Chris@10:      S *slv = (S *)X(mksolver_hc2hc)(sizeof(S), r, mkcldw);
Chris@10:      REGISTER_SOLVER(plnr, &(slv->super));
Chris@10:      if (X(mksolver_hc2hc_hook)) {
Chris@10: 	  slv = (S *)X(mksolver_hc2hc_hook)(sizeof(S), r, mkcldw);
Chris@10: 	  REGISTER_SOLVER(plnr, &(slv->super));
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: void X(hc2hc_generic_register)(planner *p)
Chris@10: {
Chris@10:      regsolver(p, 0);
Chris@10: }