root / src / grid / HPL_grid_init.c @ 1
Historique | Voir | Annoter | Télécharger (7,69 ko)
| 1 | 1 | equemene | /*
|
|---|---|---|---|
| 2 | 1 | equemene | * -- High Performance Computing Linpack Benchmark (HPL)
|
| 3 | 1 | equemene | * HPL - 2.0 - September 10, 2008
|
| 4 | 1 | equemene | * Antoine P. Petitet
|
| 5 | 1 | equemene | * University of Tennessee, Knoxville
|
| 6 | 1 | equemene | * Innovative Computing Laboratory
|
| 7 | 1 | equemene | * (C) Copyright 2000-2008 All Rights Reserved
|
| 8 | 1 | equemene | *
|
| 9 | 1 | equemene | * -- Copyright notice and Licensing terms:
|
| 10 | 1 | equemene | *
|
| 11 | 1 | equemene | * Redistribution and use in source and binary forms, with or without
|
| 12 | 1 | equemene | * modification, are permitted provided that the following conditions
|
| 13 | 1 | equemene | * are met:
|
| 14 | 1 | equemene | *
|
| 15 | 1 | equemene | * 1. Redistributions of source code must retain the above copyright
|
| 16 | 1 | equemene | * notice, this list of conditions and the following disclaimer.
|
| 17 | 1 | equemene | *
|
| 18 | 1 | equemene | * 2. Redistributions in binary form must reproduce the above copyright
|
| 19 | 1 | equemene | * notice, this list of conditions, and the following disclaimer in the
|
| 20 | 1 | equemene | * documentation and/or other materials provided with the distribution.
|
| 21 | 1 | equemene | *
|
| 22 | 1 | equemene | * 3. All advertising materials mentioning features or use of this
|
| 23 | 1 | equemene | * software must display the following acknowledgement:
|
| 24 | 1 | equemene | * This product includes software developed at the University of
|
| 25 | 1 | equemene | * Tennessee, Knoxville, Innovative Computing Laboratory.
|
| 26 | 1 | equemene | *
|
| 27 | 1 | equemene | * 4. The name of the University, the name of the Laboratory, or the
|
| 28 | 1 | equemene | * names of its contributors may not be used to endorse or promote
|
| 29 | 1 | equemene | * products derived from this software without specific written
|
| 30 | 1 | equemene | * permission.
|
| 31 | 1 | equemene | *
|
| 32 | 1 | equemene | * -- Disclaimer:
|
| 33 | 1 | equemene | *
|
| 34 | 1 | equemene | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
|
| 35 | 1 | equemene | * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
|
| 36 | 1 | equemene | * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
|
| 37 | 1 | equemene | * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY
|
| 38 | 1 | equemene | * OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
|
| 39 | 1 | equemene | * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
|
| 40 | 1 | equemene | * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
|
| 41 | 1 | equemene | * DATA OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
|
| 42 | 1 | equemene | * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
| 43 | 1 | equemene | * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
| 44 | 1 | equemene | * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
| 45 | 1 | equemene | * ---------------------------------------------------------------------
|
| 46 | 1 | equemene | */
|
| 47 | 1 | equemene | /*
|
| 48 | 1 | equemene | * Include files
|
| 49 | 1 | equemene | */
|
| 50 | 1 | equemene | #include "hpl.h" |
| 51 | 1 | equemene | |
| 52 | 1 | equemene | #ifdef STDC_HEADERS
|
| 53 | 1 | equemene | int HPL_grid_init
|
| 54 | 1 | equemene | ( |
| 55 | 1 | equemene | MPI_Comm COMM, |
| 56 | 1 | equemene | const HPL_T_ORDER ORDER,
|
| 57 | 1 | equemene | const int NPROW, |
| 58 | 1 | equemene | const int NPCOL, |
| 59 | 1 | equemene | HPL_T_grid * GRID |
| 60 | 1 | equemene | ) |
| 61 | 1 | equemene | #else
|
| 62 | 1 | equemene | int HPL_grid_init
|
| 63 | 1 | equemene | ( COMM, ORDER, NPROW, NPCOL, GRID ) |
| 64 | 1 | equemene | MPI_Comm COMM; |
| 65 | 1 | equemene | const HPL_T_ORDER ORDER;
|
| 66 | 1 | equemene | const int NPROW; |
| 67 | 1 | equemene | const int NPCOL; |
| 68 | 1 | equemene | HPL_T_grid * GRID; |
| 69 | 1 | equemene | #endif
|
| 70 | 1 | equemene | {
|
| 71 | 1 | equemene | /*
|
| 72 | 1 | equemene | * Purpose
|
| 73 | 1 | equemene | * =======
|
| 74 | 1 | equemene | *
|
| 75 | 1 | equemene | * HPL_grid_init creates a NPROW x NPCOL process grid using column- or
|
| 76 | 1 | equemene | * row-major ordering from an initial collection of processes identified
|
| 77 | 1 | equemene | * by an MPI communicator. Successful completion is indicated by the
|
| 78 | 1 | equemene | * returned error code MPI_SUCCESS. Other error codes depend on the MPI
|
| 79 | 1 | equemene | * implementation. The coordinates of processes that are not part of the
|
| 80 | 1 | equemene | * grid are set to values outside of [0..NPROW) x [0..NPCOL).
|
| 81 | 1 | equemene | *
|
| 82 | 1 | equemene | * Arguments
|
| 83 | 1 | equemene | * =========
|
| 84 | 1 | equemene | *
|
| 85 | 1 | equemene | * COMM (global/local input) MPI_Comm
|
| 86 | 1 | equemene | * On entry, COMM is the MPI communicator identifying the
|
| 87 | 1 | equemene | * initial collection of processes out of which the grid is
|
| 88 | 1 | equemene | * formed.
|
| 89 | 1 | equemene | *
|
| 90 | 1 | equemene | * ORDER (global input) const HPL_T_ORDER
|
| 91 | 1 | equemene | * On entry, ORDER specifies how the processes should be ordered
|
| 92 | 1 | equemene | * in the grid as follows:
|
| 93 | 1 | equemene | * ORDER = HPL_ROW_MAJOR row-major ordering;
|
| 94 | 1 | equemene | * ORDER = HPL_COLUMN_MAJOR column-major ordering;
|
| 95 | 1 | equemene | *
|
| 96 | 1 | equemene | * NPROW (global input) const int
|
| 97 | 1 | equemene | * On entry, NPROW specifies the number of process rows in the
|
| 98 | 1 | equemene | * grid to be created. NPROW must be at least one.
|
| 99 | 1 | equemene | *
|
| 100 | 1 | equemene | * NPCOL (global input) const int
|
| 101 | 1 | equemene | * On entry, NPCOL specifies the number of process columns in
|
| 102 | 1 | equemene | * the grid to be created. NPCOL must be at least one.
|
| 103 | 1 | equemene | *
|
| 104 | 1 | equemene | * GRID (local input/output) HPL_T_grid *
|
| 105 | 1 | equemene | * On entry, GRID points to the data structure containing the
|
| 106 | 1 | equemene | * process grid information to be initialized.
|
| 107 | 1 | equemene | *
|
| 108 | 1 | equemene | * ---------------------------------------------------------------------
|
| 109 | 1 | equemene | */
|
| 110 | 1 | equemene | /*
|
| 111 | 1 | equemene | * .. Local Variables ..
|
| 112 | 1 | equemene | */
|
| 113 | 1 | equemene | int hdim, hplerr=MPI_SUCCESS, ierr, ip2, k,
|
| 114 | 1 | equemene | mask, mycol, myrow, nprocs, rank, size; |
| 115 | 1 | equemene | /* ..
|
| 116 | 1 | equemene | * .. Executable Statements ..
|
| 117 | 1 | equemene | */
|
| 118 | 1 | equemene | MPI_Comm_rank( COMM, &rank ); MPI_Comm_size( COMM, &size ); |
| 119 | 1 | equemene | /*
|
| 120 | 1 | equemene | * Abort if illegal process grid
|
| 121 | 1 | equemene | */
|
| 122 | 1 | equemene | nprocs = NPROW * NPCOL; |
| 123 | 1 | equemene | if( ( nprocs > size ) || ( NPROW < 1 ) || ( NPCOL < 1 ) ) |
| 124 | 1 | equemene | { HPL_pabort( __LINE__, "HPL_grid_init", "Illegal Grid" ); }
|
| 125 | 1 | equemene | /*
|
| 126 | 1 | equemene | * Row- or column-major ordering of the processes
|
| 127 | 1 | equemene | */
|
| 128 | 1 | equemene | if( ORDER == HPL_ROW_MAJOR )
|
| 129 | 1 | equemene | {
|
| 130 | 1 | equemene | GRID->order = HPL_ROW_MAJOR; |
| 131 | 1 | equemene | myrow = rank / NPCOL; mycol = rank - myrow * NPCOL; |
| 132 | 1 | equemene | } |
| 133 | 1 | equemene | else
|
| 134 | 1 | equemene | {
|
| 135 | 1 | equemene | GRID->order = HPL_COLUMN_MAJOR; |
| 136 | 1 | equemene | mycol = rank / NPROW; myrow = rank - mycol * NPROW; |
| 137 | 1 | equemene | } |
| 138 | 1 | equemene | GRID->iam = rank; GRID->myrow = myrow; GRID->mycol = mycol; |
| 139 | 1 | equemene | GRID->nprow = NPROW; GRID->npcol = NPCOL; GRID->nprocs = nprocs; |
| 140 | 1 | equemene | /*
|
| 141 | 1 | equemene | * row_ip2 : largest power of two <= nprow;
|
| 142 | 1 | equemene | * row_hdim : row_ip2 procs hypercube dim;
|
| 143 | 1 | equemene | * row_ip2m1 : largest power of two <= nprow-1;
|
| 144 | 1 | equemene | * row_mask : row_ip2m1 procs hypercube mask;
|
| 145 | 1 | equemene | */
|
| 146 | 1 | equemene | hdim = 0; ip2 = 1; k = NPROW; |
| 147 | 1 | equemene | while( k > 1 ) { k >>= 1; ip2 <<= 1; hdim++; } |
| 148 | 1 | equemene | GRID->row_ip2 = ip2; GRID->row_hdim = hdim; |
| 149 | 1 | equemene | |
| 150 | 1 | equemene | mask = ip2 = 1; k = NPROW - 1; |
| 151 | 1 | equemene | while( k > 1 ) { k >>= 1; ip2 <<= 1; mask <<= 1; mask++; } |
| 152 | 1 | equemene | GRID->row_ip2m1 = ip2; GRID->row_mask = mask; |
| 153 | 1 | equemene | /*
|
| 154 | 1 | equemene | * col_ip2 : largest power of two <= npcol;
|
| 155 | 1 | equemene | * col_hdim : col_ip2 procs hypercube dim;
|
| 156 | 1 | equemene | * col_ip2m1 : largest power of two <= npcol-1;
|
| 157 | 1 | equemene | * col_mask : col_ip2m1 procs hypercube mask;
|
| 158 | 1 | equemene | */
|
| 159 | 1 | equemene | hdim = 0; ip2 = 1; k = NPCOL; |
| 160 | 1 | equemene | while( k > 1 ) { k >>= 1; ip2 <<= 1; hdim++; } |
| 161 | 1 | equemene | GRID->col_ip2 = ip2; GRID->col_hdim = hdim; |
| 162 | 1 | equemene | |
| 163 | 1 | equemene | mask = ip2 = 1; k = NPCOL - 1; |
| 164 | 1 | equemene | while( k > 1 ) { k >>= 1; ip2 <<= 1; mask <<= 1; mask++; } |
| 165 | 1 | equemene | GRID->col_ip2m1 = ip2; GRID->col_mask = mask; |
| 166 | 1 | equemene | /*
|
| 167 | 1 | equemene | * All communicator, leave if I am not part of this grid. Creation of the
|
| 168 | 1 | equemene | * row- and column communicators.
|
| 169 | 1 | equemene | */
|
| 170 | 1 | equemene | ierr = MPI_Comm_split( COMM, ( rank < nprocs ? 0 : MPI_UNDEFINED ),
|
| 171 | 1 | equemene | rank, &(GRID->all_comm) ); |
| 172 | 1 | equemene | if( GRID->all_comm == MPI_COMM_NULL ) return( ierr ); |
| 173 | 1 | equemene | |
| 174 | 1 | equemene | ierr = MPI_Comm_split( GRID->all_comm, myrow, mycol, &(GRID->row_comm) ); |
| 175 | 1 | equemene | if( ierr != MPI_SUCCESS ) hplerr = ierr;
|
| 176 | 1 | equemene | |
| 177 | 1 | equemene | ierr = MPI_Comm_split( GRID->all_comm, mycol, myrow, &(GRID->col_comm) ); |
| 178 | 1 | equemene | if( ierr != MPI_SUCCESS ) hplerr = ierr;
|
| 179 | 1 | equemene | |
| 180 | 1 | equemene | return( hplerr );
|
| 181 | 1 | equemene | /*
|
| 182 | 1 | equemene | * End of HPL_grid_init
|
| 183 | 1 | equemene | */
|
| 184 | 1 | equemene | } |