1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270
|
/*
* Copyright (C) by Argonne National Laboratory
* See COPYRIGHT in top-level directory
*/
/*
* Adapted from the example taken from MPI-The complete reference, Vol 1,
* pages 222-224.
*
* Lines after the "--CUT HERE--" were added to make this into a complete
* test program.
*/
/* Specify the maximum number of errors to report. */
#define MAX_ERRORS 10
#include "mpi.h"
#include "mpitest.h"
#include <stdio.h>
#include <stdlib.h>
#define MAX_SIZE 64
MPI_Datatype transpose_type(int M, int m, int n, MPI_Datatype type);
MPI_Datatype submatrix_type(int N, int m, int n, MPI_Datatype type);
void Transpose(float *localA, float *localB, int M, int N, MPI_Comm comm);
void Transpose(float *localA, float *localB, int M, int N, MPI_Comm comm)
/* transpose MxN matrix A that is block distributed (1-D) on
processes of comm onto block distributed matrix B */
{
int i, j, extent, myrank, p, n[2], m[2];
int lasti, lastj;
int *sendcounts, *recvcounts;
int *sdispls, *rdispls;
MPI_Datatype xtype[2][2], stype[2][2], *sendtypes, *recvtypes;
MTestPrintfMsg(2, "M = %d, N = %d\n", M, N);
/* compute parameters */
MPI_Comm_size(comm, &p);
MPI_Comm_rank(comm, &myrank);
extent = sizeof(float);
/* allocate arrays */
sendcounts = (int *) malloc(p * sizeof(int));
recvcounts = (int *) malloc(p * sizeof(int));
sdispls = (int *) malloc(p * sizeof(int));
rdispls = (int *) malloc(p * sizeof(int));
sendtypes = (MPI_Datatype *) malloc(p * sizeof(MPI_Datatype));
recvtypes = (MPI_Datatype *) malloc(p * sizeof(MPI_Datatype));
/* compute block sizes */
m[0] = M / p;
m[1] = M - (p - 1) * (M / p);
n[0] = N / p;
n[1] = N - (p - 1) * (N / p);
/* compute types */
for (i = 0; i <= 1; i++)
for (j = 0; j <= 1; j++) {
xtype[i][j] = transpose_type(N, m[i], n[j], MPI_FLOAT);
stype[i][j] = submatrix_type(M, m[i], n[j], MPI_FLOAT);
}
/* prepare collective operation arguments */
lasti = myrank == p - 1;
for (j = 0; j < p; j++) {
lastj = j == p - 1;
sendcounts[j] = 1;
sdispls[j] = j * n[0] * extent;
sendtypes[j] = xtype[lasti][lastj];
recvcounts[j] = 1;
rdispls[j] = j * m[0] * extent;
recvtypes[j] = stype[lastj][lasti];
}
/* communicate */
MTestPrintfMsg(2, "Begin persistent Alltoallw...\n");
MPI_Info info;
MPI_Info_create(&info);
MPI_Request req;
MPI_Alltoallw_init(localA, sendcounts, sdispls, sendtypes, localB, recvcounts, rdispls,
recvtypes, comm, info, &req);
for (i = 0; i < 10; ++i) {
MPI_Start(&req);
MPI_Wait(&req, MPI_STATUS_IGNORE);
}
MPI_Request_free(&req);
MPI_Info_free(&info);
MTestPrintfMsg(2, "Done with persistent Alltoallw\n");
/* Free buffers */
free(sendcounts);
free(recvcounts);
free(sdispls);
free(rdispls);
free(sendtypes);
free(recvtypes);
/* Free datatypes */
for (i = 0; i <= 1; i++)
for (j = 0; j <= 1; j++) {
MPI_Type_free(&xtype[i][j]);
MPI_Type_free(&stype[i][j]);
}
}
/* Define an n x m submatrix in a n x M local matrix (this is the
destination in the transpose matrix */
MPI_Datatype submatrix_type(int M, int m, int n, MPI_Datatype type)
/* computes a datatype for an mxn submatrix within an MxN matrix
with entries of type type */
{
/* MPI_Datatype subrow; */
MPI_Datatype submatrix;
/* The book, MPI: The Complete Reference, has the wrong type constructor
* here. Since the stride in the vector type is relative to the input
* type, the stride in the book's code is n times as long as is intended.
* Since n may not exactly divide N, it is better to simply use the
* blocklength argument in Type_vector */
/*
* MPI_Type_contiguous(n, type, &subrow);
* MPI_Type_vector(m, 1, N, subrow, &submatrix);
*/
MPI_Type_vector(n, m, M, type, &submatrix);
MPI_Type_commit(&submatrix);
/* Add a consistency test: the size of submatrix should be
* n * m * sizeof(type) and the extent should be ((n-1)*M+m) * sizeof(type) */
{
int tsize;
MPI_Aint textent, lb;
MPI_Type_size(type, &tsize);
MPI_Type_get_extent(submatrix, &lb, &textent);
if (textent != tsize * (M * (n - 1) + m)) {
fprintf(stderr, "Submatrix extent is %ld, expected %ld (%d,%d,%d)\n",
(long) textent, (long) (tsize * (M * (n - 1) + m)), M, n, m);
}
}
return (submatrix);
}
/* Extract an m x n submatrix within an m x N matrix and transpose it.
Assume storage by rows; the defined datatype accesses by columns */
MPI_Datatype transpose_type(int N, int m, int n, MPI_Datatype type)
/* computes a datatype for the transpose of an mxn matrix
with entries of type type */
{
MPI_Datatype subrow, subrow1, submatrix;
MPI_Aint lb, extent;
MPI_Type_vector(m, 1, N, type, &subrow);
MPI_Type_get_extent(type, &lb, &extent);
MPI_Type_create_resized(subrow, 0, extent, &subrow1);
MPI_Type_contiguous(n, subrow1, &submatrix);
MPI_Type_commit(&submatrix);
MPI_Type_free(&subrow);
MPI_Type_free(&subrow1);
/* Add a consistency test: the size of submatrix should be
* n * m * sizeof(type) and the extent should be ((m-1)*N+n) * sizeof(type) */
{
int tsize;
MPI_Aint textent, llb;
MPI_Type_size(type, &tsize);
MPI_Type_get_true_extent(submatrix, &llb, &textent);
if (textent != tsize * (N * (m - 1) + n)) {
fprintf(stderr, "Transpose Submatrix extent is %ld, expected %ld (%d,%d,%d)\n",
(long) textent, (long) (tsize * (N * (m - 1) + n)), N, n, m);
}
}
return (submatrix);
}
/* -- CUT HERE -- */
int main(int argc, char *argv[])
{
int gM, gN, lm, lmlast, ln, lnlast, i, j, errs = 0;
int size, rank;
float *localA, *localB;
MPI_Comm comm;
MTest_Init(&argc, &argv);
comm = MPI_COMM_WORLD;
MPI_Comm_size(comm, &size);
MPI_Comm_rank(comm, &rank);
gM = 20;
gN = 30;
/* Each block is lm x ln in size, except for the last process,
* which has lmlast x lnlast */
lm = gM / size;
lmlast = gM - (size - 1) * lm;
ln = gN / size;
lnlast = gN - (size - 1) * ln;
/* Create the local matrices.
* Initialize the input matrix so that the entries are
* consecutive integers, by row, starting at 0.
*/
if (rank == size - 1) {
localA = (float *) malloc(gN * lmlast * sizeof(float));
localB = (float *) malloc(gM * lnlast * sizeof(float));
for (i = 0; i < lmlast; i++) {
for (j = 0; j < gN; j++) {
localA[i * gN + j] = (float) (i * gN + j + rank * gN * lm);
}
}
} else {
localA = (float *) malloc(gN * lm * sizeof(float));
localB = (float *) malloc(gM * ln * sizeof(float));
for (i = 0; i < lm; i++) {
for (j = 0; j < gN; j++) {
localA[i * gN + j] = (float) (i * gN + j + rank * gN * lm);
}
}
}
MTestPrintfMsg(2, "Allocated local arrays\n");
/* Transpose */
Transpose(localA, localB, gM, gN, comm);
/* check the transposed matrix
* In the global matrix, the transpose has consecutive integers,
* organized by columns.
*/
if (rank == size - 1) {
for (i = 0; i < lnlast; i++) {
for (j = 0; j < gM; j++) {
int expected = i + gN * j + rank * ln;
if ((int) localB[i * gM + j] != expected) {
if (errs < MAX_ERRORS)
printf("Found %d but expected %d\n", (int) localB[i * gM + j], expected);
errs++;
}
}
}
} else {
for (i = 0; i < ln; i++) {
for (j = 0; j < gM; j++) {
int expected = i + gN * j + rank * ln;
if ((int) localB[i * gM + j] != expected) {
if (errs < MAX_ERRORS)
printf("Found %d but expected %d\n", (int) localB[i * gM + j], expected);
errs++;
}
}
}
}
/* Free storage */
free(localA);
free(localB);
MTest_Finalize(errs);
return MTestReturnValue(errs);
}
|