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: /* express a twiddle problem in terms of dft + multiplication by
cannam@95:    twiddle factors */
cannam@95: 
cannam@95: #include "ct.h"
cannam@95: 
cannam@95: typedef struct {
cannam@95:      ct_solver super;
cannam@95:      INT batchsz;
cannam@95: } S;
cannam@95: 
cannam@95: typedef struct {
cannam@95:      plan_dftw super;
cannam@95: 
cannam@95:      INT r, rs, m, ms, v, vs, mb, me;
cannam@95:      INT batchsz;
cannam@95:      plan *cld;
cannam@95: 
cannam@95:      triggen *t;
cannam@95:      const S *slv;
cannam@95: } P;
cannam@95: 
cannam@95: 
cannam@95: #define BATCHDIST(r) ((r) + 16)
cannam@95: 
cannam@95: /**************************************************************/
cannam@95: static void bytwiddle(const P *ego, INT mb, INT me, R *buf, R *rio, R *iio)
cannam@95: {
cannam@95:      INT j, k;
cannam@95:      INT r = ego->r, rs = ego->rs, ms = ego->ms;
cannam@95:      triggen *t = ego->t;
cannam@95:      for (j = 0; j < r; ++j) {
cannam@95: 	  for (k = mb; k < me; ++k)
cannam@95: 	       t->rotate(t, j * k,
cannam@95: 			 rio[j * rs + k * ms],
cannam@95: 			 iio[j * rs + k * ms],
cannam@95: 			 &buf[j * 2 + 2 * BATCHDIST(r) * (k - mb) + 0]);
cannam@95:      }
cannam@95: }
cannam@95: 
cannam@95: static int applicable0(const S *ego,
cannam@95: 		       INT r, INT irs, INT ors,
cannam@95: 		       INT m, INT v,
cannam@95: 		       INT mcount)
cannam@95: {
cannam@95:      return (1
cannam@95: 	     && v == 1
cannam@95: 	     && irs == ors
cannam@95: 	     && mcount >= ego->batchsz
cannam@95: 	     && mcount % ego->batchsz == 0
cannam@95: 	     && r >= 64 
cannam@95: 	     && m >= r
cannam@95: 	  );
cannam@95: }
cannam@95: 
cannam@95: static int applicable(const S *ego,
cannam@95: 		      INT r, INT irs, INT ors,
cannam@95: 		      INT m, INT v,
cannam@95: 		      INT mcount,
cannam@95: 		      const planner *plnr)
cannam@95: {
cannam@95:      if (!applicable0(ego, r, irs, ors, m, v, mcount))
cannam@95: 	  return 0;
cannam@95:      if (NO_UGLYP(plnr) && m * r < 65536)
cannam@95: 	  return 0;
cannam@95: 
cannam@95:      return 1;
cannam@95: }
cannam@95: 
cannam@95: static void dobatch(const P *ego, INT mb, INT me, R *buf, R *rio, R *iio)
cannam@95: {
cannam@95:      plan_dft *cld;
cannam@95:      INT ms = ego->ms;
cannam@95: 
cannam@95:      bytwiddle(ego, mb, me, buf, rio, iio);
cannam@95: 
cannam@95:      cld = (plan_dft *) ego->cld;
cannam@95:      cld->apply(ego->cld, buf, buf + 1, buf, buf + 1);
cannam@95:      X(cpy2d_pair_co)(buf, buf + 1,
cannam@95: 		      rio + ms * mb, iio + ms * mb,
cannam@95: 		      me-mb, 2 * BATCHDIST(ego->r), ms,
cannam@95: 		      ego->r, 2, ego->rs);
cannam@95: }
cannam@95: 
cannam@95: static void apply(const plan *ego_, R *rio, R *iio)
cannam@95: {
cannam@95:      const P *ego = (const P *) ego_;
cannam@95:      R *buf = (R *) MALLOC(sizeof(R) * 2 * BATCHDIST(ego->r) * ego->batchsz,
cannam@95: 			   BUFFERS);
cannam@95:      INT m;
cannam@95: 
cannam@95:      for (m = ego->mb; m < ego->me; m += ego->batchsz)
cannam@95: 	  dobatch(ego, m, m + ego->batchsz, buf, rio, iio);
cannam@95: 
cannam@95:      A(m == ego->me);
cannam@95: 
cannam@95:      X(ifree)(buf);
cannam@95: }
cannam@95: 
cannam@95: static void awake(plan *ego_, enum wakefulness wakefulness)
cannam@95: {
cannam@95:      P *ego = (P *) ego_;
cannam@95:      X(plan_awake)(ego->cld, wakefulness);
cannam@95: 
cannam@95:      switch (wakefulness) {
cannam@95: 	 case SLEEPY:
cannam@95: 	      X(triggen_destroy)(ego->t); ego->t = 0;
cannam@95: 	      break;
cannam@95: 	 default:
cannam@95: 	      ego->t = X(mktriggen)(AWAKE_SQRTN_TABLE, ego->r * ego->m);
cannam@95: 	      break;
cannam@95:      }
cannam@95: }
cannam@95: 
cannam@95: static void destroy(plan *ego_)
cannam@95: {
cannam@95:      P *ego = (P *) ego_;
cannam@95:      X(plan_destroy_internal)(ego->cld);
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:      p->print(p, "(dftw-genericbuf/%D-%D-%D%(%p%))",
cannam@95: 	      ego->batchsz, ego->r, ego->m, ego->cld);
cannam@95: }
cannam@95: 
cannam@95: static plan *mkcldw(const ct_solver *ego_,
cannam@95: 		    INT r, INT irs, INT ors,
cannam@95: 		    INT m, INT ms,
cannam@95: 		    INT v, INT ivs, INT ovs,
cannam@95: 		    INT mstart, INT mcount,
cannam@95: 		    R *rio, R *iio,
cannam@95: 		    planner *plnr)
cannam@95: {
cannam@95:      const S *ego = (const S *)ego_;
cannam@95:      P *pln;
cannam@95:      plan *cld = 0;
cannam@95:      R *buf;
cannam@95: 
cannam@95:      static const plan_adt padt = {
cannam@95: 	  0, awake, print, destroy
cannam@95:      };
cannam@95:      
cannam@95:      UNUSED(ivs); UNUSED(ovs); UNUSED(rio); UNUSED(iio);
cannam@95: 
cannam@95:      A(mstart >= 0 && mstart + mcount <= m);
cannam@95:      if (!applicable(ego, r, irs, ors, m, v, mcount, plnr))
cannam@95:           return (plan *)0;
cannam@95: 
cannam@95:      buf = (R *) MALLOC(sizeof(R) * 2 * BATCHDIST(r) * ego->batchsz, BUFFERS);
cannam@95:      cld = X(mkplan_d)(plnr,
cannam@95: 			X(mkproblem_dft_d)(
cannam@95: 			     X(mktensor_1d)(r, 2, 2),
cannam@95: 			     X(mktensor_1d)(ego->batchsz,
cannam@95: 					    2 * BATCHDIST(r),
cannam@95: 					    2 * BATCHDIST(r)),
cannam@95: 			     buf, buf + 1, buf, buf + 1
cannam@95: 			     )
cannam@95: 			);
cannam@95:      X(ifree)(buf);
cannam@95:      if (!cld) goto nada;
cannam@95: 
cannam@95:      pln = MKPLAN_DFTW(P, &padt, apply);
cannam@95:      pln->slv = ego;
cannam@95:      pln->cld = cld;
cannam@95:      pln->r = r;
cannam@95:      pln->m = m;
cannam@95:      pln->ms = ms;
cannam@95:      pln->rs = irs;
cannam@95:      pln->batchsz = ego->batchsz;
cannam@95:      pln->mb = mstart;
cannam@95:      pln->me = mstart + mcount;
cannam@95: 
cannam@95:      {
cannam@95: 	  double n0 = (r - 1) * (mcount - 1);
cannam@95: 	  pln->super.super.ops = cld->ops;
cannam@95: 	  pln->super.super.ops.mul += 8 * n0;
cannam@95: 	  pln->super.super.ops.add += 4 * n0;
cannam@95: 	  pln->super.super.ops.other += 8 * n0;
cannam@95:      }
cannam@95:      return &(pln->super.super);
cannam@95: 
cannam@95:  nada:
cannam@95:      X(plan_destroy_internal)(cld);
cannam@95:      return (plan *) 0;
cannam@95: }
cannam@95: 
cannam@95: static void regsolver(planner *plnr, INT r, INT batchsz)
cannam@95: {
cannam@95:      S *slv = (S *)X(mksolver_ct)(sizeof(S), r, DECDIT, mkcldw, 0);
cannam@95:      slv->batchsz = batchsz;
cannam@95:      REGISTER_SOLVER(plnr, &(slv->super.super));
cannam@95: 
cannam@95:      if (X(mksolver_ct_hook)) {
cannam@95: 	  slv = (S *)X(mksolver_ct_hook)(sizeof(S), r, DECDIT, mkcldw, 0);
cannam@95: 	  slv->batchsz = batchsz;
cannam@95: 	  REGISTER_SOLVER(plnr, &(slv->super.super));
cannam@95:      }
cannam@95: 
cannam@95: }
cannam@95: 
cannam@95: void X(ct_genericbuf_register)(planner *p)
cannam@95: {
cannam@95:      static const INT radices[] = { -1, -2, -4, -8, -16, -32, -64 };
cannam@95:      static const INT batchsizes[] = { 4, 8, 16, 32, 64 };
cannam@95:      unsigned i, j;
cannam@95: 
cannam@95:      for (i = 0; i < sizeof(radices) / sizeof(radices[0]); ++i)
cannam@95: 	  for (j = 0; j < sizeof(batchsizes) / sizeof(batchsizes[0]); ++j)
cannam@95: 	       regsolver(p, radices[i], batchsizes[j]);
cannam@95: }