1 /* Copyright (c) 2014. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
7 #include "xbt/sysdep.h"
13 typedef struct s_smpi_mpi_cart_topology {
19 } s_smpi_mpi_cart_topology_t;
21 typedef struct s_smpi_mpi_graph_topology {
26 } s_smpi_mpi_graph_topology_t;
28 typedef struct s_smpi_dist_graph_topology {
36 } s_smpi_mpi_dist_graph_topology_t;
38 typedef struct s_smpi_mpi_topology {
41 MPIR_Graph_Topology graph;
42 MPIR_Cart_Topology cart;
43 MPIR_Dist_Graph_Topology dist_graph;
45 } s_smpi_mpi_topology_t;
47 void smpi_topo_destroy(MPI_Topology topo) {
54 // smpi_graph_topo_destroy(topo->topo.graph);
57 smpi_cart_topo_destroy(topo->topo.cart);
61 // smpi_dist_graph_topo_destroy(topo->topo.dist_graph);
69 MPI_Topology smpi_topo_create(MPIR_Topo_type kind) {
70 MPI_Topology newTopo = static_cast<MPI_Topology>(xbt_malloc(sizeof(*newTopo)));
72 // Allocate and initialize the right topo should be done by the caller
76 /*******************************************************************************
77 * Cartesian topologies
78 ******************************************************************************/
79 void smpi_cart_topo_destroy(MPIR_Cart_Topology cart) {
85 xbt_free(cart->periodic);
88 xbt_free(cart->position);
94 MPI_Topology smpi_cart_topo_create(int ndims) {
95 MPI_Topology newTopo = smpi_topo_create(MPI_CART);
96 MPIR_Cart_Topology newCart = static_cast<MPIR_Cart_Topology>(xbt_malloc(sizeof(*newCart)));
98 newCart->ndims = ndims;
99 newCart->dims = xbt_new(int, ndims);
100 newCart->periodic = xbt_new(int, ndims);
101 newCart->position = xbt_new(int, ndims);
102 newTopo->topo.cart = newCart;
106 /* reorder is ignored, don't know what would be the consequences of a dumb
107 * reordering but neither do I see the point of reordering*/
108 int smpi_mpi_cart_create(MPI_Comm comm_old, int ndims, int dims[], int periods[], int reorder, MPI_Comm *comm_cart) {
109 int retval = MPI_SUCCESS;
111 MPI_Topology newCart;
112 MPI_Group newGroup, oldGroup;
113 int rank, nranks, newSize;
115 rank = smpi_comm_rank(comm_old);
119 for (i = 0 ; i < ndims ; i++) {
122 if(rank >= newSize) {
123 *comm_cart = MPI_COMM_NULL;
126 newCart = smpi_cart_topo_create(ndims);
127 oldGroup = smpi_comm_group(comm_old);
128 newGroup = smpi_group_new(newSize);
129 for (i = 0 ; i < newSize ; i++) {
130 smpi_group_set_mapping(newGroup, smpi_group_index(oldGroup, i), i);
133 newCart->topo.cart->nnodes = newSize;
135 // FIXME : code duplication... See smpi_mpi_cart_coords
137 for (i=0; i<ndims; i++) {
138 newCart->topo.cart->dims[i] = dims[i];
139 newCart->topo.cart->periodic[i] = periods[i];
140 nranks = nranks / dims[i];
141 /* FIXME: nranks could be zero (?) */
142 newCart->topo.cart->position[i] = rank / nranks;
143 rank = rank % nranks;
146 *comm_cart = smpi_comm_new(newGroup, newCart);
149 newCart = smpi_cart_topo_create(ndims);
150 *comm_cart = smpi_comm_new(smpi_group_copy(smpi_comm_group(MPI_COMM_SELF)), newCart);
152 *comm_cart = MPI_COMM_NULL;
158 int smpi_mpi_cart_sub(MPI_Comm comm, const int remain_dims[], MPI_Comm *newcomm) {
159 MPI_Topology oldTopo = smpi_comm_topo(comm);
160 int oldNDims = oldTopo->topo.cart->ndims;
161 int i, j = 0, newNDims, *newDims = NULL, *newPeriodic = NULL;
163 if (remain_dims == NULL && oldNDims != 0) {
167 for (i = 0 ; i < oldNDims ; i++) {
173 newDims = xbt_new(int, newNDims);
174 newPeriodic = xbt_new(int, newNDims);
176 // that should not segfault
177 for (i = 0 ; j < newNDims ; i++) {
179 newDims[j] = oldTopo->topo.cart->dims[i];
180 newPeriodic[j] = oldTopo->topo.cart->periodic[i];
185 return smpi_mpi_cart_create(comm, newNDims, newDims, newPeriodic, 0, newcomm);
188 int smpi_mpi_cart_coords(MPI_Comm comm, int rank, int maxdims, int coords[]) {
191 MPI_Topology topo = smpi_comm_topo(comm);
193 nnodes = topo->topo.cart->nnodes;
194 for (int i=0; i < topo->topo.cart->ndims; i++ ) {
195 nnodes = nnodes / topo->topo.cart->dims[i];
196 coords[i] = rank / nnodes;
197 rank = rank % nnodes;
202 int smpi_mpi_cart_get(MPI_Comm comm, int maxdims, int* dims, int* periods, int* coords) {
203 MPI_Topology topo = smpi_comm_topo(comm);
204 int ndims=topo->topo.cart->ndims < maxdims ? topo->topo.cart->ndims : maxdims;
205 for(int i = 0 ; i < ndims ; i++) {
206 dims[i] = topo->topo.cart->dims[i];
207 periods[i] = topo->topo.cart->periodic[i];
208 coords[i] = topo->topo.cart->position[i];
213 int smpi_mpi_cart_rank(MPI_Comm comm, int* coords, int* rank) {
214 MPI_Topology topo = smpi_comm_topo(comm);
215 int ndims = topo->topo.cart->ndims;
216 int multiplier, coord,i;
220 for ( i=ndims-1; i >=0; i-- ) {
223 /* The user can give us whatever coordinates he wants. If one of them is out of range, either this dimension is
224 * periodic, and we consider the equivalent coordinate inside the bounds, or it's not and then it's an error
226 if (coord >= topo->topo.cart->dims[i]) {
227 if ( topo->topo.cart->periodic[i] ) {
228 coord = coord % topo->topo.cart->dims[i];
230 // Should I do that ?
234 } else if (coord < 0) {
235 if(topo->topo.cart->periodic[i]) {
236 coord = coord % topo->topo.cart->dims[i];
238 coord = topo->topo.cart->dims[i] + coord;
245 *rank += multiplier * coord;
246 multiplier *= topo->topo.cart->dims[i];
251 int smpi_mpi_cart_shift(MPI_Comm comm, int direction, int disp, int *rank_source, int *rank_dest) {
252 MPI_Topology topo = smpi_comm_topo(comm);
253 int position[topo->topo.cart->ndims];
255 if(topo->topo.cart->ndims == 0) {
258 if (topo->topo.cart->ndims < direction) {
262 smpi_mpi_cart_coords(comm, smpi_comm_rank(comm), topo->topo.cart->ndims, position);
263 position[direction] += disp;
265 if(position[direction] < 0 ||
266 position[direction] >= topo->topo.cart->dims[direction]) {
267 if(topo->topo.cart->periodic[direction]) {
268 position[direction] %= topo->topo.cart->dims[direction];
269 smpi_mpi_cart_rank(comm, position, rank_dest);
271 *rank_dest = MPI_PROC_NULL;
274 smpi_mpi_cart_rank(comm, position, rank_dest);
277 position[direction] = topo->topo.cart->position[direction] - disp;
278 if(position[direction] < 0 || position[direction] >= topo->topo.cart->dims[direction]) {
279 if(topo->topo.cart->periodic[direction]) {
280 position[direction] %= topo->topo.cart->dims[direction];
281 smpi_mpi_cart_rank(comm, position, rank_source);
283 *rank_source = MPI_PROC_NULL;
286 smpi_mpi_cart_rank(comm, position, rank_source);
292 int smpi_mpi_cartdim_get(MPI_Comm comm, int *ndims) {
293 MPI_Topology topo = smpi_comm_topo(comm);
295 *ndims = topo->topo.cart->ndims;
299 // Everything below has been taken from ompi, but could be easily rewritten.
302 * Copyright (c) 2004-2007 The Trustees of Indiana University and Indiana
303 * University Research and Technology
304 * Corporation. All rights reserved.
305 * Copyright (c) 2004-2005 The University of Tennessee and The University
306 * of Tennessee Research Foundation. All rights
308 * Copyright (c) 2004-2014 High Performance Computing Center Stuttgart,
309 * University of Stuttgart. All rights reserved.
310 * Copyright (c) 2004-2005 The Regents of the University of California.
311 * All rights reserved.
312 * Copyright (c) 2012 Los Alamos National Security, LLC. All rights
314 * Copyright (c) 2014 Intel, Inc. All rights reserved
317 * Additional copyrights may follow
322 /* static functions */
323 static int assignnodes(int ndim, int nfactor, int *pfacts,int **pdims);
324 static int getfactors(int num, int *nfators, int **factors);
327 * This is a utility function, no need to have anything in the lower
328 * layer for this at all
330 int smpi_mpi_dims_create(int nnodes, int ndims, int dims[])
341 /* Get # of free-to-be-assigned processes and # of free dimensions */
344 for (i = 0, p = dims; i < ndims; ++i,++p) {
347 } else if ((*p < 0) || ((nnodes % *p) != 0)) {
356 if (freeprocs == 1) {
362 if (freeprocs == 1) {
363 for (i = 0; i < ndims; ++i, ++dims) {
371 /* Factor the number of free processes */
372 if (MPI_SUCCESS != (err = getfactors(freeprocs, &nfactors, &factors))) {
376 /* Assign free processes to free dimensions */
377 if (MPI_SUCCESS != (err = assignnodes(freedims, nfactors, factors, &procs))) {
381 /* Return assignment results */
383 for (i = 0; i < ndims; ++i, ++dims) {
389 free((char *) factors);
390 free((char *) procs);
399 * Function: - assign processes to dimensions
400 * - get "best-balanced" grid
401 * - greedy bin-packing algorithm used
402 * - sort dimensions in decreasing order
403 * - dimensions array dynamically allocated
404 * Accepts: - # of dimensions
405 * - # of prime factors
406 * - array of prime factors
407 * - ptr to array of dimensions (returned value)
408 * Returns: - 0 or ERROR
411 assignnodes(int ndim, int nfactor, int *pfacts, int **pdims)
424 /* Allocate and initialize the bins */
425 bins = (int *) malloc((unsigned) ndim * sizeof(int));
427 return MPI_ERR_NO_MEM;
431 for (i = 0, p = bins; i < ndim; ++i, ++p) {
435 /* Loop assigning factors from the highest to the lowest */
436 for (j = nfactor - 1; j >= 0; --j) {
438 /* Assign a factor to the smallest bin */
440 for (i = 1, p = pmin + 1; i < ndim; ++i, ++p) {
448 /* Sort dimensions in decreasing order (O(n^2) for now) */
449 for (i = 0, pmin = bins; i < ndim - 1; ++i, ++pmin) {
450 for (j = i + 1, p = pmin + 1; j < ndim; ++j, ++p) {
465 * Function: - factorize a number
468 * - array of prime factors
469 * Returns: - MPI_SUCCESS or ERROR
471 static int getfactors(int num, int *nfactors, int **factors) {
482 /* Allocate the array of prime factors which cannot exceed log_2(num) entries */
483 sqrtnum = ceil(sqrt(num));
484 size = ceil(log(num) / log(2));
485 *factors = (int *) malloc((unsigned) size * sizeof(int));
488 /* determine all occurences of factor 2 */
489 while((num % 2) == 0) {
493 /* determine all occurences of uneven prime numbers up to sqrt(num) */
494 for(d = 3; (num > 1) && (d < sqrtnum); d += 2) {
495 while((num % d) == 0) {
500 /* as we looped only up to sqrt(num) one factor > sqrt(num) may be left over */
502 (*factors)[i++] = num;