|
| 1 | +/* -*- Mode: C; c-basic-offset:4 ; -*- */ |
| 2 | +/* |
| 3 | + * Copyright (c) 2004-2006 The Trustees of Indiana University and Indiana |
| 4 | + * University Research and Technology |
| 5 | + * Corporation. All rights reserved. |
| 6 | + * Copyright (c) 2004-2011 The University of Tennessee and The University |
| 7 | + * of Tennessee Research Foundation. All rights |
| 8 | + * reserved. |
| 9 | + * Copyright (c) 2004-2006 High Performance Computing Center Stuttgart, |
| 10 | + * University of Stuttgart. All rights reserved. |
| 11 | + * Copyright (c) 2004-2006 The Regents of the University of California. |
| 12 | + * All rights reserved. |
| 13 | + * Copyright (c) 2009 Sun Microsystems, Inc. All rights reserved. |
| 14 | + * Copyright (c) 2009 Oak Ridge National Labs. All rights reserved. |
| 15 | + * Copyright (c) 2010 Cisco Systems, Inc. All rights reserved. |
| 16 | + * $COPYRIGHT$ |
| 17 | + * |
| 18 | + * Additional copyrights may follow |
| 19 | + * |
| 20 | + * $HEADER$ |
| 21 | + */ |
| 22 | + |
| 23 | +#include "ompi_config.h" |
| 24 | + |
| 25 | +#include <stddef.h> |
| 26 | + |
| 27 | +#include "ompi/constants.h" |
| 28 | +#include "ompi/datatype/ompi_datatype.h" |
| 29 | + |
| 30 | +static int |
| 31 | +block(const int *gsize_array, int dim, int ndims, int nprocs, |
| 32 | + int rank, int darg, int order, ptrdiff_t orig_extent, |
| 33 | + ompi_datatype_t *type_old, ompi_datatype_t **type_new, |
| 34 | + ptrdiff_t *st_offset) |
| 35 | +{ |
| 36 | + int blksize, global_size, mysize, i, j, rc, start_loop, step; |
| 37 | + ptrdiff_t stride; |
| 38 | + |
| 39 | + global_size = gsize_array[dim]; |
| 40 | + |
| 41 | + if (darg == MPI_DISTRIBUTE_DFLT_DARG) |
| 42 | + blksize = (global_size + nprocs - 1) / nprocs; |
| 43 | + else { |
| 44 | + blksize = darg; |
| 45 | + } |
| 46 | + |
| 47 | + j = global_size - blksize*rank; |
| 48 | + mysize = blksize < j ? blksize : j; |
| 49 | + if (mysize < 0) mysize = 0; |
| 50 | + |
| 51 | + if (MPI_ORDER_C == order) { |
| 52 | + start_loop = ndims - 1 ; step = -1; |
| 53 | + } else { |
| 54 | + start_loop = 0 ; step = 1; |
| 55 | + } |
| 56 | + |
| 57 | + stride = orig_extent; |
| 58 | + if (dim == start_loop) { |
| 59 | + rc = ompi_datatype_create_contiguous(mysize, type_old, type_new); |
| 60 | + if (OMPI_SUCCESS != rc) return rc; |
| 61 | + } else { |
| 62 | + for (i = start_loop ; i != dim ; i += step) { |
| 63 | + stride *= gsize_array[i]; |
| 64 | + } |
| 65 | + rc = ompi_datatype_create_hvector(mysize, 1, stride, type_old, type_new); |
| 66 | + if (OMPI_SUCCESS != rc) return rc; |
| 67 | + } |
| 68 | + |
| 69 | + *st_offset = blksize * rank; |
| 70 | + /* in terms of no. of elements of type oldtype in this dimension */ |
| 71 | + if (mysize == 0) *st_offset = 0; |
| 72 | + |
| 73 | + return OMPI_SUCCESS; |
| 74 | +} |
| 75 | + |
| 76 | + |
| 77 | +static int |
| 78 | +cyclic(const int *gsize_array, int dim, int ndims, int nprocs, |
| 79 | + int rank, int darg, int order, ptrdiff_t orig_extent, |
| 80 | + ompi_datatype_t* type_old, ompi_datatype_t **type_new, |
| 81 | + ptrdiff_t *st_offset) |
| 82 | +{ |
| 83 | + int blksize, i, blklens[2], st_index, end_index, local_size, rem, count, rc; |
| 84 | + ptrdiff_t stride, disps[2]; |
| 85 | + ompi_datatype_t *type_tmp, *types[2]; |
| 86 | + |
| 87 | + if (darg == MPI_DISTRIBUTE_DFLT_DARG) { |
| 88 | + blksize = 1; |
| 89 | + } else { |
| 90 | + blksize = darg; |
| 91 | + } |
| 92 | + |
| 93 | + st_index = rank * blksize; |
| 94 | + end_index = gsize_array[dim] - 1; |
| 95 | + |
| 96 | + if (end_index < st_index) { |
| 97 | + local_size = 0; |
| 98 | + } else { |
| 99 | + local_size = ((end_index - st_index + 1)/(nprocs*blksize))*blksize; |
| 100 | + rem = (end_index - st_index + 1) % (nprocs*blksize); |
| 101 | + local_size += rem < blksize ? rem : blksize; |
| 102 | + } |
| 103 | + |
| 104 | + count = local_size / blksize; |
| 105 | + rem = local_size % blksize; |
| 106 | + |
| 107 | + stride = nprocs*blksize*orig_extent; |
| 108 | + if (order == MPI_ORDER_FORTRAN) { |
| 109 | + for (i=0; i<dim; i++) { |
| 110 | + stride *= gsize_array[i]; |
| 111 | + } |
| 112 | + } else { |
| 113 | + for (i=ndims-1; i>dim; i--) { |
| 114 | + stride *= gsize_array[i]; |
| 115 | + } |
| 116 | + } |
| 117 | + |
| 118 | + rc = ompi_datatype_create_hvector(count, blksize, stride, type_old, type_new); |
| 119 | + if (OMPI_SUCCESS != rc) return rc; |
| 120 | + |
| 121 | + if (rem) { |
| 122 | + /* if the last block is of size less than blksize, include |
| 123 | + it separately using MPI_Type_struct */ |
| 124 | + |
| 125 | + types[0] = *type_new; |
| 126 | + types[1] = type_old; |
| 127 | + disps[0] = 0; |
| 128 | + disps[1] = count*stride; |
| 129 | + blklens[0] = 1; |
| 130 | + blklens[1] = rem; |
| 131 | + |
| 132 | + rc = ompi_datatype_create_struct(2, blklens, disps, types, &type_tmp); |
| 133 | + ompi_datatype_destroy(type_new); |
| 134 | + /* even in error condition, need to destroy type_new, so check |
| 135 | + for error after destroy. */ |
| 136 | + if (OMPI_SUCCESS != rc) return rc; |
| 137 | + *type_new = type_tmp; |
| 138 | + } |
| 139 | + |
| 140 | + /* need to set the UB for block-cyclic to work */ |
| 141 | + types[0] = *type_new; |
| 142 | + types[1] = MPI_UB; |
| 143 | + disps[0] = 0; |
| 144 | + disps[1] = orig_extent; |
| 145 | + if (order == MPI_ORDER_FORTRAN) { |
| 146 | + for (i=0; i<=dim; i++) { |
| 147 | + disps[1] *= gsize_array[i]; |
| 148 | + } |
| 149 | + } else { |
| 150 | + for (i=ndims-1; i>=dim; i--) { |
| 151 | + disps[1] *= gsize_array[i]; |
| 152 | + } |
| 153 | + } |
| 154 | + blklens[0] = blklens[1] = 1; |
| 155 | + rc = ompi_datatype_create_struct(2, blklens, disps, types, &type_tmp); |
| 156 | + ompi_datatype_destroy(type_new); |
| 157 | + /* even in error condition, need to destroy type_new, so check |
| 158 | + for error after destroy. */ |
| 159 | + if (OMPI_SUCCESS != rc) return rc; |
| 160 | + *type_new = type_tmp; |
| 161 | + |
| 162 | + *st_offset = rank * blksize; |
| 163 | + /* in terms of no. of elements of type oldtype in this dimension */ |
| 164 | + if (local_size == 0) *st_offset = 0; |
| 165 | + |
| 166 | + return OMPI_SUCCESS; |
| 167 | +} |
| 168 | + |
| 169 | +int32_t ompi_datatype_create_darray(int size, |
| 170 | + int rank, |
| 171 | + int ndims, |
| 172 | + int const* gsize_array, |
| 173 | + int const* distrib_array, |
| 174 | + int const* darg_array, |
| 175 | + int const* psize_array, |
| 176 | + int order, |
| 177 | + const ompi_datatype_t* oldtype, |
| 178 | + ompi_datatype_t** newtype) |
| 179 | +{ |
| 180 | + ompi_datatype_t *lastType; |
| 181 | + ptrdiff_t orig_extent, *st_offsets = NULL; |
| 182 | + int i, start_loop, end_loop, step; |
| 183 | + int *coords = NULL, rc = OMPI_SUCCESS; |
| 184 | + |
| 185 | + /* speedy corner case */ |
| 186 | + if (ndims < 1) { |
| 187 | + /* Don't just return MPI_DATATYPE_NULL as that can't be |
| 188 | + MPI_TYPE_FREE()ed, and that seems bad */ |
| 189 | + *newtype = ompi_datatype_create(0); |
| 190 | + ompi_datatype_add(*newtype, &ompi_mpi_datatype_null.dt, 0, 0, 0); |
| 191 | + return MPI_SUCCESS; |
| 192 | + } |
| 193 | + |
| 194 | + rc = ompi_datatype_type_extent(oldtype, &orig_extent); |
| 195 | + if (MPI_SUCCESS != rc) goto cleanup; |
| 196 | + |
| 197 | + /* calculate position in grid using row-major ordering */ |
| 198 | + { |
| 199 | + int tmp_rank = rank, procs = size; |
| 200 | + |
| 201 | + coords = (int *) malloc(ndims * sizeof(int)); |
| 202 | + for (i = 0 ; i < ndims ; i++) { |
| 203 | + procs = procs / psize_array[i]; |
| 204 | + coords[i] = tmp_rank / procs; |
| 205 | + tmp_rank = tmp_rank % procs; |
| 206 | + } |
| 207 | + } |
| 208 | + |
| 209 | + st_offsets = (ptrdiff_t *) malloc(ndims * sizeof(ptrdiff_t)); |
| 210 | + |
| 211 | + /* duplicate type to here to 1) deal with constness without |
| 212 | + casting and 2) eliminate need to for conditional destroy below. |
| 213 | + Lame, yes. But cleaner code all around. */ |
| 214 | + rc = ompi_datatype_duplicate(oldtype, &lastType); |
| 215 | + if (OMPI_SUCCESS != rc) goto cleanup; |
| 216 | + |
| 217 | + /* figure out ordering issues */ |
| 218 | + if (MPI_ORDER_C == order) { |
| 219 | + start_loop = ndims - 1 ; step = -1; end_loop = -1; |
| 220 | + } else { |
| 221 | + start_loop = 0 ; step = 1; end_loop = ndims; |
| 222 | + } |
| 223 | + |
| 224 | + /* Build up array */ |
| 225 | + for (i = start_loop ; i != end_loop; i += step) { |
| 226 | + int nprocs, tmp_rank; |
| 227 | + |
| 228 | + switch(distrib_array[i]) { |
| 229 | + case MPI_DISTRIBUTE_BLOCK: |
| 230 | + rc = block(gsize_array, i, ndims, psize_array[i], coords[i], |
| 231 | + darg_array[i], order, orig_extent, |
| 232 | + lastType, newtype, st_offsets+i); |
| 233 | + break; |
| 234 | + case MPI_DISTRIBUTE_CYCLIC: |
| 235 | + rc = cyclic(gsize_array, i, ndims, psize_array[i], coords[i], |
| 236 | + darg_array[i], order, orig_extent, |
| 237 | + lastType, newtype, st_offsets+i); |
| 238 | + break; |
| 239 | + case MPI_DISTRIBUTE_NONE: |
| 240 | + /* treat it as a block distribution on 1 process */ |
| 241 | + if (order == MPI_ORDER_C) { |
| 242 | + nprocs = psize_array[i]; tmp_rank = coords[i]; |
| 243 | + } else { |
| 244 | + nprocs = 1; tmp_rank = 0; |
| 245 | + } |
| 246 | + |
| 247 | + rc = block(gsize_array, i, ndims, nprocs, tmp_rank, |
| 248 | + MPI_DISTRIBUTE_DFLT_DARG, order, orig_extent, |
| 249 | + lastType, newtype, st_offsets+i); |
| 250 | + break; |
| 251 | + default: |
| 252 | + rc = MPI_ERR_ARG; |
| 253 | + } |
| 254 | + ompi_datatype_destroy(&lastType); |
| 255 | + /* need to destroy the old type even in error condition, so |
| 256 | + don't check return code from above until after cleanup. */ |
| 257 | + if (MPI_SUCCESS != rc) goto cleanup; |
| 258 | + lastType = *newtype; |
| 259 | + } |
| 260 | + |
| 261 | + |
| 262 | + /* set displacement and UB correctly. Use struct instead of |
| 263 | + resized for same reason as subarray */ |
| 264 | + { |
| 265 | + ptrdiff_t displs[3]; |
| 266 | + ompi_datatype_t *types[3]; |
| 267 | + int tmp_size, blength[3] = { 1, 1, 1}; |
| 268 | + |
| 269 | + displs[1] = st_offsets[start_loop]; |
| 270 | + tmp_size = 1; |
| 271 | + for (i = start_loop + step ; i != end_loop ; i += step) { |
| 272 | + tmp_size *= gsize_array[i - step]; |
| 273 | + displs[1] += tmp_size * st_offsets[i]; |
| 274 | + } |
| 275 | + |
| 276 | + displs[0] = 0; |
| 277 | + displs[1] *= orig_extent; |
| 278 | + displs[2] = orig_extent; |
| 279 | + for (i = 0 ; i < ndims ; i++) { |
| 280 | + displs[2] *= gsize_array[i]; |
| 281 | + } |
| 282 | + types[0] = MPI_LB; types[1] = lastType; types[2] = MPI_UB; |
| 283 | + |
| 284 | + rc = ompi_datatype_create_struct(3, blength, displs, types, newtype); |
| 285 | + ompi_datatype_destroy(&lastType); |
| 286 | + /* need to destroy the old type even in error condition, so |
| 287 | + don't check return code from above until after cleanup. */ |
| 288 | + if (MPI_SUCCESS != rc) goto cleanup; |
| 289 | + } |
| 290 | + |
| 291 | + cleanup: |
| 292 | + if (NULL != st_offsets) free(st_offsets); |
| 293 | + if (NULL != coords) free(coords); |
| 294 | + return OMPI_SUCCESS; |
| 295 | +} |
0 commit comments