cannam@167: /*
cannam@167:  * Copyright (c) 2003, 2007-14 Matteo Frigo
cannam@167:  * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
cannam@167:  *
cannam@167:  * This program is free software; you can redistribute it and/or modify
cannam@167:  * it under the terms of the GNU General Public License as published by
cannam@167:  * the Free Software Foundation; either version 2 of the License, or
cannam@167:  * (at your option) any later version.
cannam@167:  *
cannam@167:  * This program is distributed in the hope that it will be useful,
cannam@167:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
cannam@167:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
cannam@167:  * GNU General Public License for more details.
cannam@167:  *
cannam@167:  * You should have received a copy of the GNU General Public License
cannam@167:  * along with this program; if not, write to the Free Software
cannam@167:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
cannam@167:  *
cannam@167:  */
cannam@167: 
cannam@167: 
cannam@167: /* direct DFT solver, if we have a codelet */
cannam@167: 
cannam@167: #include "dft/dft.h"
cannam@167: 
cannam@167: typedef struct {
cannam@167:      solver super;
cannam@167:      const kdft_desc *desc;
cannam@167:      kdft k;
cannam@167:      int bufferedp;
cannam@167: } S;
cannam@167: 
cannam@167: typedef struct {
cannam@167:      plan_dft super;
cannam@167: 
cannam@167:      stride is, os, bufstride;
cannam@167:      INT n, vl, ivs, ovs;
cannam@167:      kdft k;
cannam@167:      const S *slv;
cannam@167: } P;
cannam@167: 
cannam@167: static void dobatch(const P *ego, R *ri, R *ii, R *ro, R *io, 
cannam@167: 		    R *buf, INT batchsz)
cannam@167: {
cannam@167:      X(cpy2d_pair_ci)(ri, ii, buf, buf+1,
cannam@167: 		      ego->n, WS(ego->is, 1), WS(ego->bufstride, 1),
cannam@167: 		      batchsz, ego->ivs, 2);
cannam@167:      
cannam@167:      if (IABS(WS(ego->os, 1)) < IABS(ego->ovs)) {
cannam@167: 	  /* transform directly to output */
cannam@167: 	  ego->k(buf, buf+1, ro, io, 
cannam@167: 		 ego->bufstride, ego->os, batchsz, 2, ego->ovs);
cannam@167:      } else {
cannam@167: 	  /* transform to buffer and copy back */
cannam@167: 	  ego->k(buf, buf+1, buf, buf+1, 
cannam@167: 		 ego->bufstride, ego->bufstride, batchsz, 2, 2);
cannam@167: 	  X(cpy2d_pair_co)(buf, buf+1, ro, io,
cannam@167: 			   ego->n, WS(ego->bufstride, 1), WS(ego->os, 1), 
cannam@167: 			   batchsz, 2, ego->ovs);
cannam@167:      }
cannam@167: }
cannam@167: 
cannam@167: static INT compute_batchsize(INT n)
cannam@167: {
cannam@167:      /* round up to multiple of 4 */
cannam@167:      n += 3;
cannam@167:      n &= -4;
cannam@167: 
cannam@167:      return (n + 2);
cannam@167: }
cannam@167: 
cannam@167: static void apply_buf(const plan *ego_, R *ri, R *ii, R *ro, R *io)
cannam@167: {
cannam@167:      const P *ego = (const P *) ego_;
cannam@167:      R *buf;
cannam@167:      INT vl = ego->vl, n = ego->n, batchsz = compute_batchsize(n);
cannam@167:      INT i;
cannam@167:      size_t bufsz = n * batchsz * 2 * sizeof(R);
cannam@167: 
cannam@167:      BUF_ALLOC(R *, buf, bufsz);
cannam@167: 
cannam@167:      for (i = 0; i < vl - batchsz; i += batchsz) {
cannam@167: 	  dobatch(ego, ri, ii, ro, io, buf, batchsz);
cannam@167: 	  ri += batchsz * ego->ivs; ii += batchsz * ego->ivs;
cannam@167: 	  ro += batchsz * ego->ovs; io += batchsz * ego->ovs;
cannam@167:      }
cannam@167:      dobatch(ego, ri, ii, ro, io, buf, vl - i);
cannam@167: 
cannam@167:      BUF_FREE(buf, bufsz);
cannam@167: }
cannam@167: 
cannam@167: static void apply(const plan *ego_, R *ri, R *ii, R *ro, R *io)
cannam@167: {
cannam@167:      const P *ego = (const P *) ego_;
cannam@167:      ASSERT_ALIGNED_DOUBLE;
cannam@167:      ego->k(ri, ii, ro, io, ego->is, ego->os, ego->vl, ego->ivs, ego->ovs);
cannam@167: }
cannam@167: 
cannam@167: static void apply_extra_iter(const plan *ego_, R *ri, R *ii, R *ro, R *io)
cannam@167: {
cannam@167:      const P *ego = (const P *) ego_;
cannam@167:      INT vl = ego->vl;
cannam@167: 
cannam@167:      ASSERT_ALIGNED_DOUBLE;
cannam@167: 
cannam@167:      /* for 4-way SIMD when VL is odd: iterate over an
cannam@167: 	even vector length VL, and then execute the last
cannam@167: 	iteration as a 2-vector with vector stride 0. */
cannam@167:      ego->k(ri, ii, ro, io, ego->is, ego->os, vl - 1, ego->ivs, ego->ovs);
cannam@167: 
cannam@167:      ego->k(ri + (vl - 1) * ego->ivs, ii + (vl - 1) * ego->ivs,
cannam@167: 	    ro + (vl - 1) * ego->ovs, io + (vl - 1) * ego->ovs,
cannam@167: 	    ego->is, ego->os, 1, 0, 0);
cannam@167: }
cannam@167: 
cannam@167: static void destroy(plan *ego_)
cannam@167: {
cannam@167:      P *ego = (P *) ego_;
cannam@167:      X(stride_destroy)(ego->is);
cannam@167:      X(stride_destroy)(ego->os);
cannam@167:      X(stride_destroy)(ego->bufstride);
cannam@167: }
cannam@167: 
cannam@167: static void print(const plan *ego_, printer *p)
cannam@167: {
cannam@167:      const P *ego = (const P *) ego_;
cannam@167:      const S *s = ego->slv;
cannam@167:      const kdft_desc *d = s->desc;
cannam@167: 
cannam@167:      if (ego->slv->bufferedp)
cannam@167: 	  p->print(p, "(dft-directbuf/%D-%D%v \"%s\")", 
cannam@167: 		   compute_batchsize(d->sz), d->sz, ego->vl, d->nam);
cannam@167:      else
cannam@167: 	  p->print(p, "(dft-direct-%D%v \"%s\")", d->sz, ego->vl, d->nam);
cannam@167: }
cannam@167: 
cannam@167: static int applicable_buf(const solver *ego_, const problem *p_,
cannam@167: 			  const planner *plnr)
cannam@167: {
cannam@167:      const S *ego = (const S *) ego_;
cannam@167:      const problem_dft *p = (const problem_dft *) p_;
cannam@167:      const kdft_desc *d = ego->desc;
cannam@167:      INT vl;
cannam@167:      INT ivs, ovs;
cannam@167:      INT batchsz;
cannam@167: 
cannam@167:      return (
cannam@167: 	  1
cannam@167: 	  && p->sz->rnk == 1
cannam@167: 	  && p->vecsz->rnk == 1
cannam@167: 	  && p->sz->dims[0].n == d->sz
cannam@167: 
cannam@167: 	  /* check strides etc */
cannam@167: 	  && X(tensor_tornk1)(p->vecsz, &vl, &ivs, &ovs)
cannam@167: 
cannam@167: 	  /* UGLY if IS <= IVS */
cannam@167: 	  && !(NO_UGLYP(plnr) &&
cannam@167: 	       X(iabs)(p->sz->dims[0].is) <= X(iabs)(ivs))
cannam@167: 
cannam@167: 	  && (batchsz = compute_batchsize(d->sz), 1)
cannam@167: 	  && (d->genus->okp(d, 0, ((const R *)0) + 1, p->ro, p->io,
cannam@167: 			    2 * batchsz, p->sz->dims[0].os,
cannam@167: 			    batchsz, 2, ovs, plnr))
cannam@167: 	  && (d->genus->okp(d, 0, ((const R *)0) + 1, p->ro, p->io,
cannam@167: 			    2 * batchsz, p->sz->dims[0].os,
cannam@167: 			    vl % batchsz, 2, ovs, plnr))
cannam@167: 
cannam@167: 
cannam@167: 	  && (0
cannam@167: 	      /* can operate out-of-place */
cannam@167: 	      || p->ri != p->ro
cannam@167: 
cannam@167: 	      /* can operate in-place as long as strides are the same */
cannam@167: 	      || X(tensor_inplace_strides2)(p->sz, p->vecsz)
cannam@167: 
cannam@167: 	      /* can do it if the problem fits in the buffer, no matter
cannam@167: 		 what the strides are */
cannam@167: 	      || vl <= batchsz
cannam@167: 	       )
cannam@167: 	  );
cannam@167: }
cannam@167: 
cannam@167: static int applicable(const solver *ego_, const problem *p_,
cannam@167: 		      const planner *plnr, int *extra_iterp)
cannam@167: {
cannam@167:      const S *ego = (const S *) ego_;
cannam@167:      const problem_dft *p = (const problem_dft *) p_;
cannam@167:      const kdft_desc *d = ego->desc;
cannam@167:      INT vl;
cannam@167:      INT ivs, ovs;
cannam@167: 
cannam@167:      return (
cannam@167: 	  1
cannam@167: 	  && p->sz->rnk == 1
cannam@167: 	  && p->vecsz->rnk <= 1
cannam@167: 	  && p->sz->dims[0].n == d->sz
cannam@167: 
cannam@167: 	  /* check strides etc */
cannam@167: 	  && X(tensor_tornk1)(p->vecsz, &vl, &ivs, &ovs)
cannam@167: 
cannam@167: 	  && ((*extra_iterp = 0,
cannam@167: 	       (d->genus->okp(d, p->ri, p->ii, p->ro, p->io,
cannam@167: 			      p->sz->dims[0].is, p->sz->dims[0].os,
cannam@167: 			      vl, ivs, ovs, plnr)))
cannam@167: 	      ||
cannam@167: 	      (*extra_iterp = 1,
cannam@167: 	       ((d->genus->okp(d, p->ri, p->ii, p->ro, p->io,
cannam@167: 			       p->sz->dims[0].is, p->sz->dims[0].os,
cannam@167: 			       vl - 1, ivs, ovs, plnr))
cannam@167: 		&&
cannam@167: 		(d->genus->okp(d, p->ri, p->ii, p->ro, p->io,
cannam@167: 			       p->sz->dims[0].is, p->sz->dims[0].os,
cannam@167: 			       2, 0, 0, plnr)))))
cannam@167: 
cannam@167: 	  && (0
cannam@167: 	      /* can operate out-of-place */
cannam@167: 	      || p->ri != p->ro
cannam@167: 
cannam@167: 	      /* can always compute one transform */
cannam@167: 	      || vl == 1
cannam@167: 
cannam@167: 	      /* can operate in-place as long as strides are the same */
cannam@167: 	      || X(tensor_inplace_strides2)(p->sz, p->vecsz)
cannam@167: 	       )
cannam@167: 	  );
cannam@167: }
cannam@167: 
cannam@167: 
cannam@167: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
cannam@167: {
cannam@167:      const S *ego = (const S *) ego_;
cannam@167:      P *pln;
cannam@167:      const problem_dft *p;
cannam@167:      iodim *d;
cannam@167:      const kdft_desc *e = ego->desc;
cannam@167: 
cannam@167:      static const plan_adt padt = {
cannam@167: 	  X(dft_solve), X(null_awake), print, destroy
cannam@167:      };
cannam@167: 
cannam@167:      UNUSED(plnr);
cannam@167: 
cannam@167:      if (ego->bufferedp) {
cannam@167: 	  if (!applicable_buf(ego_, p_, plnr))
cannam@167: 	       return (plan *)0;
cannam@167: 	  pln = MKPLAN_DFT(P, &padt, apply_buf);
cannam@167:      } else {
cannam@167: 	  int extra_iterp = 0;
cannam@167: 	  if (!applicable(ego_, p_, plnr, &extra_iterp))
cannam@167: 	       return (plan *)0;
cannam@167: 	  pln = MKPLAN_DFT(P, &padt, extra_iterp ? apply_extra_iter : apply);
cannam@167:      }
cannam@167: 
cannam@167:      p = (const problem_dft *) p_;
cannam@167:      d = p->sz->dims;
cannam@167:      pln->k = ego->k;
cannam@167:      pln->n = d[0].n;
cannam@167:      pln->is = X(mkstride)(pln->n, d[0].is);
cannam@167:      pln->os = X(mkstride)(pln->n, d[0].os);
cannam@167:      pln->bufstride = X(mkstride)(pln->n, 2 * compute_batchsize(pln->n));
cannam@167: 
cannam@167:      X(tensor_tornk1)(p->vecsz, &pln->vl, &pln->ivs, &pln->ovs);
cannam@167:      pln->slv = ego;
cannam@167: 
cannam@167:      X(ops_zero)(&pln->super.super.ops);
cannam@167:      X(ops_madd2)(pln->vl / e->genus->vl, &e->ops, &pln->super.super.ops);
cannam@167: 
cannam@167:      if (ego->bufferedp) 
cannam@167: 	  pln->super.super.ops.other += 4 * pln->n * pln->vl;
cannam@167: 
cannam@167:      pln->super.super.could_prune_now_p = !ego->bufferedp;
cannam@167:      return &(pln->super.super);
cannam@167: }
cannam@167: 
cannam@167: static solver *mksolver(kdft k, const kdft_desc *desc, int bufferedp)
cannam@167: {
cannam@167:      static const solver_adt sadt = { PROBLEM_DFT, mkplan, 0 };
cannam@167:      S *slv = MKSOLVER(S, &sadt);
cannam@167:      slv->k = k;
cannam@167:      slv->desc = desc;
cannam@167:      slv->bufferedp = bufferedp;
cannam@167:      return &(slv->super);
cannam@167: }
cannam@167: 
cannam@167: solver *X(mksolver_dft_direct)(kdft k, const kdft_desc *desc)
cannam@167: {
cannam@167:      return mksolver(k, desc, 0);
cannam@167: }
cannam@167: 
cannam@167: solver *X(mksolver_dft_directbuf)(kdft k, const kdft_desc *desc)
cannam@167: {
cannam@167:      return mksolver(k, desc, 1);
cannam@167: }