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: /* plans for distributed out-of-place transpose using MPI_Alltoall,
cannam@95:    and which destroy the input array (unless TRANSPOSED_IN is used) */
cannam@95: 
cannam@95: #include "mpi-transpose.h"
cannam@95: #include <string.h>
cannam@95: 
cannam@95: typedef struct {
cannam@95:      solver super;
cannam@95:      int copy_transposed_in; /* whether to copy the input for TRANSPOSED_IN,
cannam@95: 				which makes the final transpose out-of-place
cannam@95: 				but costs an extra copy and requires us
cannam@95: 				to destroy the input */
cannam@95: } S;
cannam@95: 
cannam@95: typedef struct {
cannam@95:      plan_mpi_transpose super;
cannam@95: 
cannam@95:      plan *cld1, *cld2, *cld2rest, *cld3;
cannam@95: 
cannam@95:      MPI_Comm comm;
cannam@95:      int *send_block_sizes, *send_block_offsets;
cannam@95:      int *recv_block_sizes, *recv_block_offsets;
cannam@95: 
cannam@95:      INT rest_Ioff, rest_Ooff;
cannam@95: 
cannam@95:      int equal_blocks;
cannam@95: } P;
cannam@95: 
cannam@95: static void apply(const plan *ego_, R *I, R *O)
cannam@95: {
cannam@95:      const P *ego = (const P *) ego_;
cannam@95:      plan_rdft *cld1, *cld2, *cld2rest, *cld3;
cannam@95: 
cannam@95:      /* transpose locally to get contiguous chunks */
cannam@95:      cld1 = (plan_rdft *) ego->cld1;
cannam@95:      if (cld1) {
cannam@95: 	  cld1->apply(ego->cld1, I, O);
cannam@95: 	  
cannam@95: 	  /* transpose chunks globally */
cannam@95: 	  if (ego->equal_blocks)
cannam@95: 	       MPI_Alltoall(O, ego->send_block_sizes[0], FFTW_MPI_TYPE,
cannam@95: 			    I, ego->recv_block_sizes[0], FFTW_MPI_TYPE,
cannam@95: 			    ego->comm);
cannam@95: 	  else
cannam@95: 	       MPI_Alltoallv(O, ego->send_block_sizes, ego->send_block_offsets,
cannam@95: 			     FFTW_MPI_TYPE,
cannam@95: 			     I, ego->recv_block_sizes, ego->recv_block_offsets,
cannam@95: 			     FFTW_MPI_TYPE,
cannam@95: 			     ego->comm);
cannam@95:      }
cannam@95:      else { /* TRANSPOSED_IN, no need to destroy input */
cannam@95: 	  /* transpose chunks globally */
cannam@95: 	  if (ego->equal_blocks)
cannam@95: 	       MPI_Alltoall(I, ego->send_block_sizes[0], FFTW_MPI_TYPE,
cannam@95: 			    O, ego->recv_block_sizes[0], FFTW_MPI_TYPE,
cannam@95: 			    ego->comm);
cannam@95: 	  else
cannam@95: 	       MPI_Alltoallv(I, ego->send_block_sizes, ego->send_block_offsets,
cannam@95: 			     FFTW_MPI_TYPE,
cannam@95: 			     O, ego->recv_block_sizes, ego->recv_block_offsets,
cannam@95: 			     FFTW_MPI_TYPE,
cannam@95: 			     ego->comm);
cannam@95: 	  I = O; /* final transpose (if any) is in-place */
cannam@95:      }
cannam@95:      
cannam@95:      /* transpose locally, again, to get ordinary row-major */
cannam@95:      cld2 = (plan_rdft *) ego->cld2;
cannam@95:      if (cld2) {
cannam@95: 	  cld2->apply(ego->cld2, I, O);
cannam@95: 	  cld2rest = (plan_rdft *) ego->cld2rest;
cannam@95: 	  if (cld2rest) { /* leftover from unequal block sizes */
cannam@95: 	       cld2rest->apply(ego->cld2rest,
cannam@95: 			       I + ego->rest_Ioff, O + ego->rest_Ooff);
cannam@95: 	       cld3 = (plan_rdft *) ego->cld3;
cannam@95: 	       if (cld3)
cannam@95: 		    cld3->apply(ego->cld3, O, O);
cannam@95: 	       /* else TRANSPOSED_OUT is true and user wants O transposed */
cannam@95: 	  }
cannam@95:      }
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_mpi_transpose *p = (const problem_mpi_transpose *) p_;
cannam@95:      return (1
cannam@95: 	     && p->I != p->O
cannam@95: 	     && (!NO_DESTROY_INPUTP(plnr) || 
cannam@95: 		 ((p->flags & TRANSPOSED_IN) && !ego->copy_transposed_in))
cannam@95: 	     && ((p->flags & TRANSPOSED_IN) || !ego->copy_transposed_in)
cannam@95: 	     && ONLY_TRANSPOSEDP(p->flags)
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:      X(plan_awake)(ego->cld1, wakefulness);
cannam@95:      X(plan_awake)(ego->cld2, wakefulness);
cannam@95:      X(plan_awake)(ego->cld2rest, wakefulness);
cannam@95:      X(plan_awake)(ego->cld3, wakefulness);
cannam@95: }
cannam@95: 
cannam@95: static void destroy(plan *ego_)
cannam@95: {
cannam@95:      P *ego = (P *) ego_;
cannam@95:      X(ifree0)(ego->send_block_sizes);
cannam@95:      MPI_Comm_free(&ego->comm);
cannam@95:      X(plan_destroy_internal)(ego->cld3);
cannam@95:      X(plan_destroy_internal)(ego->cld2rest);
cannam@95:      X(plan_destroy_internal)(ego->cld2);
cannam@95:      X(plan_destroy_internal)(ego->cld1);
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, "(mpi-transpose-alltoall%s%(%p%)%(%p%)%(%p%)%(%p%))",
cannam@95: 	      ego->equal_blocks ? "/e" : "",
cannam@95: 	      ego->cld1, ego->cld2, ego->cld2rest, ego->cld3);
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_mpi_transpose *p;
cannam@95:      P *pln;
cannam@95:      plan *cld1 = 0, *cld2 = 0, *cld2rest = 0, *cld3 = 0;
cannam@95:      INT b, bt, vn, rest_Ioff, rest_Ooff;
cannam@95:      R *I;
cannam@95:      int *sbs, *sbo, *rbs, *rbo;
cannam@95:      int pe, my_pe, n_pes;
cannam@95:      int equal_blocks = 1;
cannam@95:      static const plan_adt padt = {
cannam@95:           XM(transpose_solve), awake, print, destroy
cannam@95:      };
cannam@95: 
cannam@95:      if (!applicable(ego, p_, plnr))
cannam@95:           return (plan *) 0;
cannam@95: 
cannam@95:      p = (const problem_mpi_transpose *) p_;
cannam@95:      vn = p->vn;
cannam@95: 
cannam@95:      MPI_Comm_rank(p->comm, &my_pe);
cannam@95:      MPI_Comm_size(p->comm, &n_pes);
cannam@95: 
cannam@95:      b = XM(block)(p->nx, p->block, my_pe);
cannam@95: 
cannam@95:      if (p->flags & TRANSPOSED_IN) { /* I is already transposed */
cannam@95: 	  if (ego->copy_transposed_in) {
cannam@95: 	       cld1 = X(mkplan_f_d)(plnr,
cannam@95: 				  X(mkproblem_rdft_0_d)(X(mktensor_1d)
cannam@95: 							(b * p->ny * vn, 1, 1),
cannam@95: 							I = p->I, p->O),
cannam@95: 				    0, 0, NO_SLOW);
cannam@95: 	       if (XM(any_true)(!cld1, p->comm)) goto nada;
cannam@95: 	  }
cannam@95: 	  else
cannam@95: 	       I = p->O; /* final transpose is in-place */
cannam@95:      }
cannam@95:      else { /* transpose b x ny x vn -> ny x b x vn */
cannam@95: 	  cld1 = X(mkplan_f_d)(plnr, 
cannam@95: 			       X(mkproblem_rdft_0_d)(X(mktensor_3d)
cannam@95: 						     (b, p->ny * vn, vn,
cannam@95: 						      p->ny, vn, b * vn,
cannam@95: 						      vn, 1, 1),
cannam@95: 						     I = p->I, p->O),
cannam@95: 			       0, 0, NO_SLOW);
cannam@95: 	  if (XM(any_true)(!cld1, p->comm)) goto nada;
cannam@95:      }
cannam@95: 	  
cannam@95:      if (XM(any_true)(!XM(mkplans_posttranspose)(p, plnr, I, p->O, my_pe,
cannam@95: 						 &cld2, &cld2rest, &cld3,
cannam@95: 						 &rest_Ioff, &rest_Ooff),
cannam@95: 		      p->comm)) goto nada;
cannam@95: 
cannam@95:      pln = MKPLAN_MPI_TRANSPOSE(P, &padt, apply);
cannam@95: 
cannam@95:      pln->cld1 = cld1;
cannam@95:      pln->cld2 = cld2;
cannam@95:      pln->cld2rest = cld2rest;
cannam@95:      pln->rest_Ioff = rest_Ioff;
cannam@95:      pln->rest_Ooff = rest_Ooff;
cannam@95:      pln->cld3 = cld3;
cannam@95: 
cannam@95:      MPI_Comm_dup(p->comm, &pln->comm);
cannam@95: 
cannam@95:      /* Compute sizes/offsets of blocks to send for all-to-all command. */
cannam@95:      sbs = (int *) MALLOC(4 * n_pes * sizeof(int), PLANS);
cannam@95:      sbo = sbs + n_pes;
cannam@95:      rbs = sbo + n_pes;
cannam@95:      rbo = rbs + n_pes;
cannam@95:      b = XM(block)(p->nx, p->block, my_pe);
cannam@95:      bt = XM(block)(p->ny, p->tblock, my_pe);
cannam@95:      for (pe = 0; pe < n_pes; ++pe) {
cannam@95: 	  INT db, dbt; /* destination block sizes */
cannam@95: 	  db = XM(block)(p->nx, p->block, pe);
cannam@95: 	  dbt = XM(block)(p->ny, p->tblock, pe);
cannam@95: 	  if (db != p->block || dbt != p->tblock)
cannam@95: 	       equal_blocks = 0;
cannam@95: 
cannam@95: 	  /* MPI requires type "int" here; apparently it
cannam@95: 	     has no 64-bit API?  Grrr. */
cannam@95: 	  sbs[pe] = (int) (b * dbt * vn);
cannam@95: 	  sbo[pe] = (int) (pe * (b * p->tblock) * vn);
cannam@95: 	  rbs[pe] = (int) (db * bt * vn);
cannam@95: 	  rbo[pe] = (int) (pe * (p->block * bt) * vn);
cannam@95:      }
cannam@95:      pln->send_block_sizes = sbs;
cannam@95:      pln->send_block_offsets = sbo;
cannam@95:      pln->recv_block_sizes = rbs;
cannam@95:      pln->recv_block_offsets = rbo;
cannam@95:      pln->equal_blocks = equal_blocks;
cannam@95: 
cannam@95:      X(ops_zero)(&pln->super.super.ops);
cannam@95:      if (cld1) X(ops_add2)(&cld1->ops, &pln->super.super.ops);
cannam@95:      if (cld2) X(ops_add2)(&cld2->ops, &pln->super.super.ops);
cannam@95:      if (cld2rest) X(ops_add2)(&cld2rest->ops, &pln->super.super.ops);
cannam@95:      if (cld3) X(ops_add2)(&cld3->ops, &pln->super.super.ops);
cannam@95:      /* FIXME: should MPI operations be counted in "other" somehow? */
cannam@95: 
cannam@95:      return &(pln->super.super);
cannam@95: 
cannam@95:  nada:
cannam@95:      X(plan_destroy_internal)(cld3);
cannam@95:      X(plan_destroy_internal)(cld2rest);
cannam@95:      X(plan_destroy_internal)(cld2);
cannam@95:      X(plan_destroy_internal)(cld1);
cannam@95:      return (plan *) 0;
cannam@95: }
cannam@95: 
cannam@95: static solver *mksolver(int copy_transposed_in)
cannam@95: {
cannam@95:      static const solver_adt sadt = { PROBLEM_MPI_TRANSPOSE, mkplan, 0 };
cannam@95:      S *slv = MKSOLVER(S, &sadt);
cannam@95:      slv->copy_transposed_in = copy_transposed_in;
cannam@95:      return &(slv->super);
cannam@95: }
cannam@95: 
cannam@95: void XM(transpose_alltoall_register)(planner *p)
cannam@95: {
cannam@95:      int cti;
cannam@95:      for (cti = 0; cti <= 1; ++cti)
cannam@95: 	  REGISTER_SOLVER(p, mksolver(cti));
cannam@95: }