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: /* Complex DFTs of rank == 1 via six-step algorithm. */
Chris@10: 
Chris@10: #include "mpi-dft.h"
Chris@10: #include "mpi-transpose.h"
Chris@10: #include "dft.h"
Chris@10: 
Chris@10: typedef struct {
Chris@10:      solver super;
Chris@10:      rdftapply apply; /* apply_ddft_first or apply_ddft_last */
Chris@10:      int preserve_input; /* preserve input even if DESTROY_INPUT was passed */
Chris@10: } S;
Chris@10: 
Chris@10: typedef struct {
Chris@10:      plan_mpi_dft super;
Chris@10: 
Chris@10:      triggen *t;
Chris@10:      plan *cldt, *cld_ddft, *cld_dft;
Chris@10:      INT roff, ioff;
Chris@10:      int preserve_input;
Chris@10:      INT vn, xmin, xmax, xs, m, r;
Chris@10: } P;
Chris@10: 
Chris@10: static void do_twiddle(triggen *t, INT ir, INT m, INT vn, R *xr, R *xi)
Chris@10: {
Chris@10:      void (*rotate)(triggen *, INT, R, R, R *) = t->rotate;
Chris@10:      INT im, iv;
Chris@10:      for (im = 0; im < m; ++im)
Chris@10: 	  for (iv = 0; iv < vn; ++iv) {
Chris@10: 	       /* TODO: modify/inline rotate function
Chris@10: 		  so that it can do whole vn vector at once? */
Chris@10: 	       R c[2];
Chris@10: 	       rotate(t, ir * im, *xr, *xi, c);
Chris@10: 	       *xr = c[0]; *xi = c[1];
Chris@10: 	       xr += 2; xi += 2;
Chris@10: 	  }
Chris@10: }
Chris@10: 
Chris@10: /* radix-r DFT of size r*m.  This is equivalent to an m x r 2d DFT,
Chris@10:    plus twiddle factors between the size-m and size-r 1d DFTs, where
Chris@10:    the m dimension is initially distributed.  The output is transposed
Chris@10:    to r x m where the r dimension is distributed. 
Chris@10: 
Chris@10:    This algorithm follows the general sequence:
Chris@10:         global transpose (m x r -> r x m)
Chris@10:         DFTs of size m
Chris@10: 	multiply by twiddles + global transpose (r x m -> m x r)
Chris@10: 	DFTs of size r
Chris@10: 	global transpose (m x r -> r x m)
Chris@10:    where the multiplication by twiddles can come before or after
Chris@10:    the middle transpose.  The first/last transposes are omitted
Chris@10:    for SCRAMBLED_IN/OUT formats, respectively.
Chris@10: 
Chris@10:    However, we wish to exploit our dft-rank1-bigvec solver, which
Chris@10:    solves a vector of distributed DFTs via transpose+dft+transpose.
Chris@10:    Therefore, we can group *either* the DFTs of size m *or* the
Chris@10:    DFTs of size r with their surrounding transposes as a single
Chris@10:    distributed-DFT (ddft) plan.  These two variations correspond to
Chris@10:    apply_ddft_first or apply_ddft_last, respectively.
Chris@10: */
Chris@10: 
Chris@10: static void apply_ddft_first(const plan *ego_, R *I, R *O)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      plan_dft *cld_dft;
Chris@10:      plan_rdft *cldt, *cld_ddft;
Chris@10:      INT roff, ioff, im, mmax, ms, r, vn;
Chris@10:      triggen *t;
Chris@10:      R *dI, *dO;
Chris@10: 
Chris@10:      /* distributed size-m DFTs, with output in m x r format */
Chris@10:      cld_ddft = (plan_rdft *) ego->cld_ddft;
Chris@10:      cld_ddft->apply(ego->cld_ddft, I, O);
Chris@10: 
Chris@10:      cldt = (plan_rdft *) ego->cldt;
Chris@10:      if (ego->preserve_input || !cldt) I = O;
Chris@10: 
Chris@10:      /* twiddle multiplications, followed by 1d DFTs of size-r */
Chris@10:      cld_dft = (plan_dft *) ego->cld_dft;
Chris@10:      roff = ego->roff; ioff = ego->ioff;
Chris@10:      mmax = ego->xmax; ms = ego->xs;
Chris@10:      t = ego->t; r = ego->r; vn = ego->vn;
Chris@10:      dI = O; dO = I;
Chris@10:      for (im = ego->xmin; im <= mmax; ++im) {
Chris@10: 	  do_twiddle(t, im, r, vn, dI+roff, dI+ioff);
Chris@10: 	  cld_dft->apply((plan *) cld_dft, dI+roff, dI+ioff, dO+roff, dO+ioff);
Chris@10: 	  dI += ms; dO += ms;
Chris@10:      }
Chris@10: 
Chris@10:      /* final global transpose (m x r -> r x m), if not SCRAMBLED_OUT */
Chris@10:      if (cldt) 
Chris@10: 	  cldt->apply((plan *) cldt, I, O);
Chris@10: }
Chris@10: 
Chris@10: static void apply_ddft_last(const plan *ego_, R *I, R *O)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      plan_dft *cld_dft;
Chris@10:      plan_rdft *cldt, *cld_ddft;
Chris@10:      INT roff, ioff, ir, rmax, rs, m, vn;
Chris@10:      triggen *t;
Chris@10:      R *dI, *dO0, *dO;
Chris@10: 
Chris@10:      /* initial global transpose (m x r -> r x m), if not SCRAMBLED_IN */
Chris@10:      cldt = (plan_rdft *) ego->cldt;
Chris@10:      if (cldt) {
Chris@10: 	  cldt->apply((plan *) cldt, I, O);
Chris@10: 	  dI = O;
Chris@10:      }
Chris@10:      else 
Chris@10: 	  dI = I;
Chris@10:      if (ego->preserve_input) dO = O; else dO = I;
Chris@10:      dO0 = dO;
Chris@10: 
Chris@10:      /* 1d DFTs of size m, followed by twiddle multiplications */
Chris@10:      cld_dft = (plan_dft *) ego->cld_dft;
Chris@10:      roff = ego->roff; ioff = ego->ioff;
Chris@10:      rmax = ego->xmax; rs = ego->xs;
Chris@10:      t = ego->t; m = ego->m; vn = ego->vn;
Chris@10:      for (ir = ego->xmin; ir <= rmax; ++ir) {
Chris@10: 	  cld_dft->apply((plan *) cld_dft, dI+roff, dI+ioff, dO+roff, dO+ioff);
Chris@10: 	  do_twiddle(t, ir, m, vn, dO+roff, dO+ioff);
Chris@10: 	  dI += rs; dO += rs;
Chris@10:      }
Chris@10: 
Chris@10:      /* distributed size-r DFTs, with output in r x m format */
Chris@10:      cld_ddft = (plan_rdft *) ego->cld_ddft;
Chris@10:      cld_ddft->apply(ego->cld_ddft, dO0, O);
Chris@10: }
Chris@10: 
Chris@10: static int applicable(const S *ego, const problem *p_,
Chris@10: 		      const planner *plnr,
Chris@10: 		      INT *r, INT rblock[2], INT mblock[2])
Chris@10: {
Chris@10:      const problem_mpi_dft *p = (const problem_mpi_dft *) p_;
Chris@10:      int n_pes;
Chris@10:      MPI_Comm_size(p->comm, &n_pes);
Chris@10:      return (1
Chris@10: 	     && p->sz->rnk == 1
Chris@10: 
Chris@10: 	     && ONLY_SCRAMBLEDP(p->flags)
Chris@10: 
Chris@10: 	     && (!ego->preserve_input || (!NO_DESTROY_INPUTP(plnr)
Chris@10:                                           && p->I != p->O))
Chris@10: 
Chris@10: 	     && (!(p->flags & SCRAMBLED_IN) || ego->apply == apply_ddft_last)
Chris@10: 	     && (!(p->flags & SCRAMBLED_OUT) || ego->apply == apply_ddft_first)
Chris@10: 
Chris@10: 	     && (!NO_SLOWP(plnr) /* slow if dft-serial is applicable */
Chris@10:                  || !XM(dft_serial_applicable)(p))
Chris@10: 
Chris@10: 	     /* disallow if dft-rank1-bigvec is applicable since the
Chris@10: 		data distribution may be slightly different (ugh!) */
Chris@10: 	     && (p->vn < n_pes || p->flags)
Chris@10: 
Chris@10: 	     && (*r = XM(choose_radix)(p->sz->dims[0], n_pes,
Chris@10: 				       p->flags, p->sign,
Chris@10: 				       rblock, mblock))
Chris@10: 
Chris@10: 	     /* ddft_first or last has substantial advantages in the
Chris@10: 		bigvec transpositions for the common case where
Chris@10: 		n_pes == n/r or r, respectively */
Chris@10: 	     && (!NO_UGLYP(plnr)
Chris@10: 		 || !(*r == n_pes && ego->apply == apply_ddft_first)
Chris@10: 		 || !(p->sz->dims[0].n / *r == n_pes 
Chris@10: 		      && ego->apply == apply_ddft_last))
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->cldt, wakefulness);
Chris@10:      X(plan_awake)(ego->cld_dft, wakefulness);
Chris@10:      X(plan_awake)(ego->cld_ddft, wakefulness);
Chris@10: 
Chris@10:      switch (wakefulness) {
Chris@10:          case SLEEPY:
Chris@10:               X(triggen_destroy)(ego->t); ego->t = 0;
Chris@10:               break;
Chris@10:          default:
Chris@10:               ego->t = X(mktriggen)(AWAKE_SQRTN_TABLE, ego->r * ego->m);
Chris@10:               break;
Chris@10:      }
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->cldt);
Chris@10:      X(plan_destroy_internal)(ego->cld_dft);
Chris@10:      X(plan_destroy_internal)(ego->cld_ddft);
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, "(mpi-dft-rank1/%D%s%s%(%p%)%(%p%)%(%p%))",
Chris@10: 	      ego->r,
Chris@10: 	      ego->super.apply == apply_ddft_first ? "/first" : "/last",
Chris@10: 	      ego->preserve_input==2 ?"/p":"",
Chris@10: 	      ego->cld_ddft, ego->cld_dft, ego->cldt);
Chris@10: }
Chris@10: 
Chris@10: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
Chris@10: {
Chris@10:      const S *ego = (const S *) ego_;
Chris@10:      const problem_mpi_dft *p;
Chris@10:      P *pln;
Chris@10:      plan *cld_dft = 0, *cld_ddft = 0, *cldt = 0;
Chris@10:      R *ri, *ii, *ro, *io, *I, *O;
Chris@10:      INT r, rblock[2], m, mblock[2], rp, mp, mpblock[2], mpb;
Chris@10:      int my_pe, n_pes, preserve_input, ddft_first;
Chris@10:      dtensor *sz;
Chris@10:      static const plan_adt padt = {
Chris@10:           XM(dft_solve), awake, print, destroy
Chris@10:      };
Chris@10: 
Chris@10:      UNUSED(ego);
Chris@10: 
Chris@10:      if (!applicable(ego, p_, plnr, &r, rblock, mblock))
Chris@10:           return (plan *) 0;
Chris@10: 
Chris@10:      p = (const problem_mpi_dft *) p_;
Chris@10: 
Chris@10:      MPI_Comm_rank(p->comm, &my_pe);
Chris@10:      MPI_Comm_size(p->comm, &n_pes);
Chris@10: 
Chris@10:      m = p->sz->dims[0].n / r;
Chris@10: 
Chris@10:      /* some hackery so that we can plan both ddft_first and ddft_last
Chris@10: 	as if they were ddft_first */
Chris@10:      if ((ddft_first = (ego->apply == apply_ddft_first))) {
Chris@10: 	  rp = r; mp = m;
Chris@10: 	  mpblock[IB] = mblock[IB]; mpblock[OB] = mblock[OB];
Chris@10: 	  mpb = XM(block)(mp, mpblock[OB], my_pe);
Chris@10:      }
Chris@10:      else {
Chris@10: 	  rp = m; mp = r;
Chris@10: 	  mpblock[IB] = rblock[IB]; mpblock[OB] = rblock[OB];
Chris@10: 	  mpb = XM(block)(mp, mpblock[IB], my_pe);
Chris@10:      }
Chris@10: 
Chris@10:      preserve_input = ego->preserve_input ? 2 : NO_DESTROY_INPUTP(plnr);
Chris@10: 
Chris@10:      sz = XM(mkdtensor)(1);
Chris@10:      sz->dims[0].n = mp;
Chris@10:      sz->dims[0].b[IB] = mpblock[IB];
Chris@10:      sz->dims[0].b[OB] = mpblock[OB];
Chris@10:      I = (ddft_first || !preserve_input) ? p->I : p->O;
Chris@10:      O = p->O;
Chris@10:      cld_ddft = X(mkplan_d)(plnr, XM(mkproblem_dft_d)(sz, rp * p->vn,
Chris@10: 						      I, O, p->comm, p->sign,
Chris@10: 						      RANK1_BIGVEC_ONLY));
Chris@10:      if (XM(any_true)(!cld_ddft, p->comm)) goto nada;
Chris@10: 
Chris@10:      I = TAINT((ddft_first || !p->flags) ? p->O : p->I, rp * p->vn * 2);
Chris@10:      O = TAINT((preserve_input || (ddft_first && p->flags)) ? p->O : p->I, 
Chris@10: 	       rp * p->vn * 2);
Chris@10:      X(extract_reim)(p->sign, I, &ri, &ii);
Chris@10:      X(extract_reim)(p->sign, O, &ro, &io);
Chris@10:      cld_dft = X(mkplan_d)(plnr,
Chris@10: 			X(mkproblem_dft_d)(X(mktensor_1d)(rp, p->vn*2,p->vn*2),
Chris@10: 					   X(mktensor_1d)(p->vn, 2, 2),
Chris@10: 					   ri, ii, ro, io));
Chris@10:      if (XM(any_true)(!cld_dft, p->comm)) goto nada;
Chris@10:      
Chris@10:      if (!p->flags) { /* !(SCRAMBLED_IN or SCRAMBLED_OUT) */
Chris@10: 	  I = (ddft_first && preserve_input) ? p->O : p->I;
Chris@10: 	  O = p->O;
Chris@10: 	  cldt = X(mkplan_d)(plnr,
Chris@10: 			     XM(mkproblem_transpose)(
Chris@10: 				  m, r, p->vn * 2,
Chris@10: 				  I, O,
Chris@10: 				  ddft_first ? mblock[OB] : mblock[IB],
Chris@10: 				  ddft_first ? rblock[OB] : rblock[IB],
Chris@10: 				  p->comm, 0));
Chris@10: 	  if (XM(any_true)(!cldt, p->comm)) goto nada;	  
Chris@10:      }
Chris@10: 
Chris@10:      pln = MKPLAN_MPI_DFT(P, &padt, ego->apply);
Chris@10: 
Chris@10:      pln->cld_ddft = cld_ddft;
Chris@10:      pln->cld_dft = cld_dft;
Chris@10:      pln->cldt = cldt;
Chris@10:      pln->preserve_input = preserve_input;
Chris@10:      X(extract_reim)(p->sign, p->O, &ro, &io);
Chris@10:      pln->roff = ro - p->O;
Chris@10:      pln->ioff = io - p->O;
Chris@10:      pln->vn = p->vn;
Chris@10:      pln->m = m;
Chris@10:      pln->r = r;
Chris@10:      pln->xmin = (ddft_first ? mblock[OB] : rblock[IB]) * my_pe;
Chris@10:      pln->xmax = pln->xmin + mpb - 1;
Chris@10:      pln->xs = rp * p->vn * 2;
Chris@10:      pln->t = 0;
Chris@10: 
Chris@10:      X(ops_add)(&cld_ddft->ops, &cld_dft->ops, &pln->super.super.ops);
Chris@10:      if (cldt) X(ops_add2)(&cldt->ops, &pln->super.super.ops);
Chris@10:      {
Chris@10:           double n0 = (1 + pln->xmax - pln->xmin) * (mp - 1) * pln->vn;
Chris@10:           pln->super.super.ops.mul += 8 * n0;
Chris@10:           pln->super.super.ops.add += 4 * n0;
Chris@10:           pln->super.super.ops.other += 8 * n0;
Chris@10:      }
Chris@10: 
Chris@10:      return &(pln->super.super);
Chris@10: 
Chris@10:  nada:
Chris@10:      X(plan_destroy_internal)(cldt);
Chris@10:      X(plan_destroy_internal)(cld_dft);
Chris@10:      X(plan_destroy_internal)(cld_ddft);
Chris@10:      return (plan *) 0;
Chris@10: }
Chris@10: 
Chris@10: static solver *mksolver(rdftapply apply, int preserve_input)
Chris@10: {
Chris@10:      static const solver_adt sadt = { PROBLEM_MPI_DFT, mkplan, 0 };
Chris@10:      S *slv = MKSOLVER(S, &sadt);
Chris@10:      slv->apply = apply;
Chris@10:      slv->preserve_input = preserve_input;
Chris@10:      return &(slv->super);
Chris@10: }
Chris@10: 
Chris@10: void XM(dft_rank1_register)(planner *p)
Chris@10: {
Chris@10:      rdftapply apply[] = { apply_ddft_first, apply_ddft_last };
Chris@10:      unsigned int iapply;
Chris@10:      int preserve_input;
Chris@10:      for (iapply = 0; iapply < sizeof(apply) / sizeof(apply[0]); ++iapply)
Chris@10: 	  for (preserve_input = 0; preserve_input <= 1; ++preserve_input)
Chris@10: 	       REGISTER_SOLVER(p, mksolver(apply[iapply], preserve_input));
Chris@10: }