root / src / pgesv / HPL_rollT.c @ 8
Historique | Voir | Annoter | Télécharger (10,03 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 | #define I_SEND 0 |
| 53 | 1 | equemene | #define I_RECV 1 |
| 54 | 1 | equemene | |
| 55 | 1 | equemene | #ifdef STDC_HEADERS
|
| 56 | 1 | equemene | void HPL_rollT
|
| 57 | 1 | equemene | ( |
| 58 | 1 | equemene | HPL_T_panel * PBCST, |
| 59 | 1 | equemene | int * IFLAG,
|
| 60 | 1 | equemene | HPL_T_panel * PANEL, |
| 61 | 1 | equemene | const int N, |
| 62 | 1 | equemene | double * U,
|
| 63 | 1 | equemene | const int LDU, |
| 64 | 1 | equemene | const int * IPLEN, |
| 65 | 1 | equemene | const int * IPMAP, |
| 66 | 1 | equemene | const int * IPMAPM1 |
| 67 | 1 | equemene | ) |
| 68 | 1 | equemene | #else
|
| 69 | 1 | equemene | void HPL_rollT
|
| 70 | 1 | equemene | ( PBCST, IFLAG, PANEL, N, U, LDU, IPLEN, IPMAP, IPMAPM1 ) |
| 71 | 1 | equemene | HPL_T_panel * PBCST; |
| 72 | 1 | equemene | int * IFLAG;
|
| 73 | 1 | equemene | HPL_T_panel * PANEL; |
| 74 | 1 | equemene | const int N; |
| 75 | 1 | equemene | double * U;
|
| 76 | 1 | equemene | const int LDU; |
| 77 | 1 | equemene | const int * IPLEN; |
| 78 | 1 | equemene | const int * IPMAP; |
| 79 | 1 | equemene | const int * IPMAPM1; |
| 80 | 1 | equemene | #endif
|
| 81 | 1 | equemene | {
|
| 82 | 1 | equemene | /*
|
| 83 | 1 | equemene | * Purpose
|
| 84 | 1 | equemene | * =======
|
| 85 | 1 | equemene | *
|
| 86 | 1 | equemene | * HPL_rollT rolls the local arrays containing the local pieces of U, so
|
| 87 | 1 | equemene | * that on exit to this function U is replicated in every process row.
|
| 88 | 1 | equemene | * In addition, this function probe for the presence of the column panel
|
| 89 | 1 | equemene | * and forwards it when available.
|
| 90 | 1 | equemene | *
|
| 91 | 1 | equemene | * Arguments
|
| 92 | 1 | equemene | * =========
|
| 93 | 1 | equemene | *
|
| 94 | 1 | equemene | * PBCST (local input/output) HPL_T_panel *
|
| 95 | 1 | equemene | * On entry, PBCST points to the data structure containing the
|
| 96 | 1 | equemene | * panel (to be broadcast) information.
|
| 97 | 1 | equemene | *
|
| 98 | 1 | equemene | * IFLAG (local input/output) int *
|
| 99 | 1 | equemene | * On entry, IFLAG indicates whether or not the broadcast has
|
| 100 | 1 | equemene | * already been completed. If not, probing will occur, and the
|
| 101 | 1 | equemene | * outcome will be contained in IFLAG on exit.
|
| 102 | 1 | equemene | *
|
| 103 | 1 | equemene | * PANEL (local input/output) HPL_T_panel *
|
| 104 | 1 | equemene | * On entry, PANEL points to the data structure containing the
|
| 105 | 1 | equemene | * panel (to be rolled) information.
|
| 106 | 1 | equemene | *
|
| 107 | 1 | equemene | * N (local input) const int
|
| 108 | 1 | equemene | * On entry, N specifies the local number of rows of U. N must
|
| 109 | 1 | equemene | * be at least zero.
|
| 110 | 1 | equemene | *
|
| 111 | 1 | equemene | * U (local input/output) double *
|
| 112 | 1 | equemene | * On entry, U is an array of dimension (LDU,*) containing the
|
| 113 | 1 | equemene | * local pieces of U in each process row.
|
| 114 | 1 | equemene | *
|
| 115 | 1 | equemene | * LDU (local input) const int
|
| 116 | 1 | equemene | * On entry, LDU specifies the local leading dimension of U. LDU
|
| 117 | 1 | equemene | * should be at least MAX(1,N).
|
| 118 | 1 | equemene | *
|
| 119 | 1 | equemene | * IPLEN (global input) const int *
|
| 120 | 1 | equemene | * On entry, IPLEN is an array of dimension NPROW+1. This array
|
| 121 | 1 | equemene | * is such that IPLEN[i+1] - IPLEN[i] is the number of rows of U
|
| 122 | 1 | equemene | * in each process row.
|
| 123 | 1 | equemene | *
|
| 124 | 1 | equemene | * IPMAP (global input) const int *
|
| 125 | 1 | equemene | * On entry, IMAP is an array of dimension NPROW. This array
|
| 126 | 1 | equemene | * contains the logarithmic mapping of the processes. In other
|
| 127 | 1 | equemene | * words, IMAP[myrow] is the absolute coordinate of the sorted
|
| 128 | 1 | equemene | * process.
|
| 129 | 1 | equemene | *
|
| 130 | 1 | equemene | * IPMAPM1 (global input) const int *
|
| 131 | 1 | equemene | * On entry, IMAPM1 is an array of dimension NPROW. This array
|
| 132 | 1 | equemene | * contains the inverse of the logarithmic mapping contained in
|
| 133 | 1 | equemene | * IMAP: For i in [0.. NPROW) IMAPM1[IMAP[i]] = i.
|
| 134 | 1 | equemene | *
|
| 135 | 1 | equemene | * ---------------------------------------------------------------------
|
| 136 | 1 | equemene | */
|
| 137 | 1 | equemene | /*
|
| 138 | 1 | equemene | * .. Local Variables ..
|
| 139 | 1 | equemene | */
|
| 140 | 1 | equemene | #if 0
|
| 141 | 1 | equemene | MPI_Datatype type[2];
|
| 142 | 1 | equemene | #endif
|
| 143 | 1 | equemene | MPI_Status status; |
| 144 | 1 | equemene | MPI_Request request; |
| 145 | 1 | equemene | MPI_Comm comm; |
| 146 | 1 | equemene | int Cmsgid=MSGID_BEGIN_PFACT, ibufR, ibufS,
|
| 147 | 1 | equemene | ierr=MPI_SUCCESS, il, k, l, lengthR, |
| 148 | 1 | equemene | lengthS, mydist, myrow, next, npm1, nprow, |
| 149 | 1 | equemene | partner, prev; |
| 150 | 1 | equemene | /* ..
|
| 151 | 1 | equemene | * .. Executable Statements ..
|
| 152 | 1 | equemene | */
|
| 153 | 1 | equemene | if( N <= 0 ) return; |
| 154 | 1 | equemene | |
| 155 | 1 | equemene | npm1 = ( nprow = PANEL->grid->nprow ) - 1; myrow = PANEL->grid->myrow;
|
| 156 | 1 | equemene | comm = PANEL->grid->col_comm; |
| 157 | 1 | equemene | /*
|
| 158 | 1 | equemene | * Rolling phase
|
| 159 | 1 | equemene | */
|
| 160 | 1 | equemene | mydist = IPMAPM1[myrow]; |
| 161 | 1 | equemene | prev = IPMAP[MModSub1( mydist, nprow )]; |
| 162 | 1 | equemene | next = IPMAP[MModAdd1( mydist, nprow )]; |
| 163 | 1 | equemene | |
| 164 | 1 | equemene | for( k = 0; k < npm1; k++ ) |
| 165 | 1 | equemene | {
|
| 166 | 1 | equemene | l = (int)( (unsigned int)(k) >> 1 ); |
| 167 | 1 | equemene | |
| 168 | 1 | equemene | if( ( ( mydist + k ) & 1 ) != 0 ) |
| 169 | 1 | equemene | {
|
| 170 | 1 | equemene | il = MModAdd( mydist, l, nprow ); |
| 171 | 1 | equemene | lengthS = IPLEN[il+1] - ( ibufS = IPLEN[il] );
|
| 172 | 1 | equemene | il = MModSub( mydist, l+1, nprow );
|
| 173 | 1 | equemene | lengthR = IPLEN[il+1] - ( ibufR = IPLEN[il] ); partner = prev;
|
| 174 | 1 | equemene | } |
| 175 | 1 | equemene | else
|
| 176 | 1 | equemene | {
|
| 177 | 1 | equemene | il = MModSub( mydist, l, nprow ); |
| 178 | 1 | equemene | lengthS = IPLEN[il+1] - ( ibufS = IPLEN[il] );
|
| 179 | 1 | equemene | il = MModAdd( mydist, l+1, nprow );
|
| 180 | 1 | equemene | lengthR = IPLEN[il+1] - ( ibufR = IPLEN[il] ); partner = next;
|
| 181 | 1 | equemene | } |
| 182 | 1 | equemene | |
| 183 | 1 | equemene | if( lengthR > 0 ) |
| 184 | 1 | equemene | {
|
| 185 | 1 | equemene | #if 0
|
| 186 | 1 | equemene | if( ierr == MPI_SUCCESS )
|
| 187 | 1 | equemene | {
|
| 188 | 1 | equemene | if( LDU == N )
|
| 189 | 1 | equemene | ierr = MPI_Type_contiguous( lengthR * LDU, MPI_DOUBLE,
|
| 190 | 1 | equemene | &type[I_RECV] );
|
| 191 | 1 | equemene | else
|
| 192 | 1 | equemene | ierr = MPI_Type_vector( lengthR, N, LDU, MPI_DOUBLE,
|
| 193 | 1 | equemene | &type[I_RECV] );
|
| 194 | 1 | equemene | }
|
| 195 | 1 | equemene | if( ierr == MPI_SUCCESS )
|
| 196 | 1 | equemene | ierr = MPI_Type_commit( &type[I_RECV] );
|
| 197 | 1 | equemene | if( ierr == MPI_SUCCESS )
|
| 198 | 1 | equemene | ierr = MPI_Irecv( Mptr( U, 0, ibufR, LDU ), 1, type[I_RECV],
|
| 199 | 1 | equemene | partner, Cmsgid, comm, &request );
|
| 200 | 1 | equemene | #else
|
| 201 | 1 | equemene | /*
|
| 202 | 1 | equemene | * In our case, LDU is N - Do not use the MPI datatype.
|
| 203 | 1 | equemene | */
|
| 204 | 1 | equemene | if( ierr == MPI_SUCCESS )
|
| 205 | 1 | equemene | ierr = MPI_Irecv( Mptr( U, 0, ibufR, LDU ), lengthR*LDU,
|
| 206 | 1 | equemene | MPI_DOUBLE, partner, Cmsgid, comm, &request ); |
| 207 | 1 | equemene | #endif
|
| 208 | 1 | equemene | } |
| 209 | 1 | equemene | |
| 210 | 1 | equemene | if( lengthS > 0 ) |
| 211 | 1 | equemene | {
|
| 212 | 1 | equemene | #if 0
|
| 213 | 1 | equemene | if( ierr == MPI_SUCCESS )
|
| 214 | 1 | equemene | {
|
| 215 | 1 | equemene | if( LDU == N )
|
| 216 | 1 | equemene | ierr = MPI_Type_contiguous( lengthS*LDU, MPI_DOUBLE,
|
| 217 | 1 | equemene | &type[I_SEND] );
|
| 218 | 1 | equemene | else
|
| 219 | 1 | equemene | ierr = MPI_Type_vector( lengthS, N, LDU, MPI_DOUBLE,
|
| 220 | 1 | equemene | &type[I_SEND] );
|
| 221 | 1 | equemene | }
|
| 222 | 1 | equemene | if( ierr == MPI_SUCCESS )
|
| 223 | 1 | equemene | ierr = MPI_Type_commit( &type[I_SEND] );
|
| 224 | 1 | equemene | if( ierr == MPI_SUCCESS )
|
| 225 | 1 | equemene | ierr = MPI_Send( Mptr( U, 0, ibufS, LDU ), 1, type[I_SEND],
|
| 226 | 1 | equemene | partner, Cmsgid, comm );
|
| 227 | 1 | equemene | if( ierr == MPI_SUCCESS )
|
| 228 | 1 | equemene | ierr = MPI_Type_free( &type[I_SEND] );
|
| 229 | 1 | equemene | #else
|
| 230 | 1 | equemene | /*
|
| 231 | 1 | equemene | * In our case, LDU is N - Do not use the MPI datatype.
|
| 232 | 1 | equemene | */
|
| 233 | 1 | equemene | if( ierr == MPI_SUCCESS )
|
| 234 | 1 | equemene | ierr = MPI_Send( Mptr( U, 0, ibufS, LDU ), lengthS*LDU,
|
| 235 | 1 | equemene | MPI_DOUBLE, partner, Cmsgid, comm ); |
| 236 | 1 | equemene | #endif
|
| 237 | 1 | equemene | } |
| 238 | 1 | equemene | |
| 239 | 1 | equemene | if( lengthR > 0 ) |
| 240 | 1 | equemene | {
|
| 241 | 1 | equemene | if( ierr == MPI_SUCCESS )
|
| 242 | 1 | equemene | ierr = MPI_Wait( &request, &status ); |
| 243 | 1 | equemene | #if 0
|
| 244 | 1 | equemene | if( ierr == MPI_SUCCESS )
|
| 245 | 1 | equemene | ierr = MPI_Type_free( &type[I_RECV] );
|
| 246 | 1 | equemene | #endif
|
| 247 | 1 | equemene | } |
| 248 | 1 | equemene | /*
|
| 249 | 1 | equemene | * Probe for column panel - forward it when available
|
| 250 | 1 | equemene | */
|
| 251 | 1 | equemene | if( *IFLAG == HPL_KEEP_TESTING ) (void) HPL_bcast( PBCST, IFLAG ); |
| 252 | 1 | equemene | } |
| 253 | 1 | equemene | |
| 254 | 1 | equemene | if( ierr != MPI_SUCCESS )
|
| 255 | 1 | equemene | { HPL_pabort( __LINE__, "HPL_rollT", "MPI call failed" ); }
|
| 256 | 1 | equemene | /*
|
| 257 | 1 | equemene | * End of HPL_rollT
|
| 258 | 1 | equemene | */
|
| 259 | 1 | equemene | } |