root / src / pgesv / HPL_rollT.c
Historique | Voir | Annoter | Télécharger (10,03 ko)
| 1 |
/*
|
|---|---|
| 2 |
* -- High Performance Computing Linpack Benchmark (HPL)
|
| 3 |
* HPL - 2.0 - September 10, 2008
|
| 4 |
* Antoine P. Petitet
|
| 5 |
* University of Tennessee, Knoxville
|
| 6 |
* Innovative Computing Laboratory
|
| 7 |
* (C) Copyright 2000-2008 All Rights Reserved
|
| 8 |
*
|
| 9 |
* -- Copyright notice and Licensing terms:
|
| 10 |
*
|
| 11 |
* Redistribution and use in source and binary forms, with or without
|
| 12 |
* modification, are permitted provided that the following conditions
|
| 13 |
* are met:
|
| 14 |
*
|
| 15 |
* 1. Redistributions of source code must retain the above copyright
|
| 16 |
* notice, this list of conditions and the following disclaimer.
|
| 17 |
*
|
| 18 |
* 2. Redistributions in binary form must reproduce the above copyright
|
| 19 |
* notice, this list of conditions, and the following disclaimer in the
|
| 20 |
* documentation and/or other materials provided with the distribution.
|
| 21 |
*
|
| 22 |
* 3. All advertising materials mentioning features or use of this
|
| 23 |
* software must display the following acknowledgement:
|
| 24 |
* This product includes software developed at the University of
|
| 25 |
* Tennessee, Knoxville, Innovative Computing Laboratory.
|
| 26 |
*
|
| 27 |
* 4. The name of the University, the name of the Laboratory, or the
|
| 28 |
* names of its contributors may not be used to endorse or promote
|
| 29 |
* products derived from this software without specific written
|
| 30 |
* permission.
|
| 31 |
*
|
| 32 |
* -- Disclaimer:
|
| 33 |
*
|
| 34 |
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
|
| 35 |
* ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
|
| 36 |
* LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
|
| 37 |
* A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY
|
| 38 |
* OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
|
| 39 |
* SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
|
| 40 |
* LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
|
| 41 |
* DATA OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
|
| 42 |
* THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
| 43 |
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
| 44 |
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
| 45 |
* ---------------------------------------------------------------------
|
| 46 |
*/
|
| 47 |
/*
|
| 48 |
* Include files
|
| 49 |
*/
|
| 50 |
#include "hpl.h" |
| 51 |
|
| 52 |
#define I_SEND 0 |
| 53 |
#define I_RECV 1 |
| 54 |
|
| 55 |
#ifdef STDC_HEADERS
|
| 56 |
void HPL_rollT
|
| 57 |
( |
| 58 |
HPL_T_panel * PBCST, |
| 59 |
int * IFLAG,
|
| 60 |
HPL_T_panel * PANEL, |
| 61 |
const int N, |
| 62 |
double * U,
|
| 63 |
const int LDU, |
| 64 |
const int * IPLEN, |
| 65 |
const int * IPMAP, |
| 66 |
const int * IPMAPM1 |
| 67 |
) |
| 68 |
#else
|
| 69 |
void HPL_rollT
|
| 70 |
( PBCST, IFLAG, PANEL, N, U, LDU, IPLEN, IPMAP, IPMAPM1 ) |
| 71 |
HPL_T_panel * PBCST; |
| 72 |
int * IFLAG;
|
| 73 |
HPL_T_panel * PANEL; |
| 74 |
const int N; |
| 75 |
double * U;
|
| 76 |
const int LDU; |
| 77 |
const int * IPLEN; |
| 78 |
const int * IPMAP; |
| 79 |
const int * IPMAPM1; |
| 80 |
#endif
|
| 81 |
{
|
| 82 |
/*
|
| 83 |
* Purpose
|
| 84 |
* =======
|
| 85 |
*
|
| 86 |
* HPL_rollT rolls the local arrays containing the local pieces of U, so
|
| 87 |
* that on exit to this function U is replicated in every process row.
|
| 88 |
* In addition, this function probe for the presence of the column panel
|
| 89 |
* and forwards it when available.
|
| 90 |
*
|
| 91 |
* Arguments
|
| 92 |
* =========
|
| 93 |
*
|
| 94 |
* PBCST (local input/output) HPL_T_panel *
|
| 95 |
* On entry, PBCST points to the data structure containing the
|
| 96 |
* panel (to be broadcast) information.
|
| 97 |
*
|
| 98 |
* IFLAG (local input/output) int *
|
| 99 |
* On entry, IFLAG indicates whether or not the broadcast has
|
| 100 |
* already been completed. If not, probing will occur, and the
|
| 101 |
* outcome will be contained in IFLAG on exit.
|
| 102 |
*
|
| 103 |
* PANEL (local input/output) HPL_T_panel *
|
| 104 |
* On entry, PANEL points to the data structure containing the
|
| 105 |
* panel (to be rolled) information.
|
| 106 |
*
|
| 107 |
* N (local input) const int
|
| 108 |
* On entry, N specifies the local number of rows of U. N must
|
| 109 |
* be at least zero.
|
| 110 |
*
|
| 111 |
* U (local input/output) double *
|
| 112 |
* On entry, U is an array of dimension (LDU,*) containing the
|
| 113 |
* local pieces of U in each process row.
|
| 114 |
*
|
| 115 |
* LDU (local input) const int
|
| 116 |
* On entry, LDU specifies the local leading dimension of U. LDU
|
| 117 |
* should be at least MAX(1,N).
|
| 118 |
*
|
| 119 |
* IPLEN (global input) const int *
|
| 120 |
* On entry, IPLEN is an array of dimension NPROW+1. This array
|
| 121 |
* is such that IPLEN[i+1] - IPLEN[i] is the number of rows of U
|
| 122 |
* in each process row.
|
| 123 |
*
|
| 124 |
* IPMAP (global input) const int *
|
| 125 |
* On entry, IMAP is an array of dimension NPROW. This array
|
| 126 |
* contains the logarithmic mapping of the processes. In other
|
| 127 |
* words, IMAP[myrow] is the absolute coordinate of the sorted
|
| 128 |
* process.
|
| 129 |
*
|
| 130 |
* IPMAPM1 (global input) const int *
|
| 131 |
* On entry, IMAPM1 is an array of dimension NPROW. This array
|
| 132 |
* contains the inverse of the logarithmic mapping contained in
|
| 133 |
* IMAP: For i in [0.. NPROW) IMAPM1[IMAP[i]] = i.
|
| 134 |
*
|
| 135 |
* ---------------------------------------------------------------------
|
| 136 |
*/
|
| 137 |
/*
|
| 138 |
* .. Local Variables ..
|
| 139 |
*/
|
| 140 |
#if 0
|
| 141 |
MPI_Datatype type[2];
|
| 142 |
#endif
|
| 143 |
MPI_Status status; |
| 144 |
MPI_Request request; |
| 145 |
MPI_Comm comm; |
| 146 |
int Cmsgid=MSGID_BEGIN_PFACT, ibufR, ibufS,
|
| 147 |
ierr=MPI_SUCCESS, il, k, l, lengthR, |
| 148 |
lengthS, mydist, myrow, next, npm1, nprow, |
| 149 |
partner, prev; |
| 150 |
/* ..
|
| 151 |
* .. Executable Statements ..
|
| 152 |
*/
|
| 153 |
if( N <= 0 ) return; |
| 154 |
|
| 155 |
npm1 = ( nprow = PANEL->grid->nprow ) - 1; myrow = PANEL->grid->myrow;
|
| 156 |
comm = PANEL->grid->col_comm; |
| 157 |
/*
|
| 158 |
* Rolling phase
|
| 159 |
*/
|
| 160 |
mydist = IPMAPM1[myrow]; |
| 161 |
prev = IPMAP[MModSub1( mydist, nprow )]; |
| 162 |
next = IPMAP[MModAdd1( mydist, nprow )]; |
| 163 |
|
| 164 |
for( k = 0; k < npm1; k++ ) |
| 165 |
{
|
| 166 |
l = (int)( (unsigned int)(k) >> 1 ); |
| 167 |
|
| 168 |
if( ( ( mydist + k ) & 1 ) != 0 ) |
| 169 |
{
|
| 170 |
il = MModAdd( mydist, l, nprow ); |
| 171 |
lengthS = IPLEN[il+1] - ( ibufS = IPLEN[il] );
|
| 172 |
il = MModSub( mydist, l+1, nprow );
|
| 173 |
lengthR = IPLEN[il+1] - ( ibufR = IPLEN[il] ); partner = prev;
|
| 174 |
} |
| 175 |
else
|
| 176 |
{
|
| 177 |
il = MModSub( mydist, l, nprow ); |
| 178 |
lengthS = IPLEN[il+1] - ( ibufS = IPLEN[il] );
|
| 179 |
il = MModAdd( mydist, l+1, nprow );
|
| 180 |
lengthR = IPLEN[il+1] - ( ibufR = IPLEN[il] ); partner = next;
|
| 181 |
} |
| 182 |
|
| 183 |
if( lengthR > 0 ) |
| 184 |
{
|
| 185 |
#if 0
|
| 186 |
if( ierr == MPI_SUCCESS )
|
| 187 |
{
|
| 188 |
if( LDU == N )
|
| 189 |
ierr = MPI_Type_contiguous( lengthR * LDU, MPI_DOUBLE,
|
| 190 |
&type[I_RECV] );
|
| 191 |
else
|
| 192 |
ierr = MPI_Type_vector( lengthR, N, LDU, MPI_DOUBLE,
|
| 193 |
&type[I_RECV] );
|
| 194 |
}
|
| 195 |
if( ierr == MPI_SUCCESS )
|
| 196 |
ierr = MPI_Type_commit( &type[I_RECV] );
|
| 197 |
if( ierr == MPI_SUCCESS )
|
| 198 |
ierr = MPI_Irecv( Mptr( U, 0, ibufR, LDU ), 1, type[I_RECV],
|
| 199 |
partner, Cmsgid, comm, &request );
|
| 200 |
#else
|
| 201 |
/*
|
| 202 |
* In our case, LDU is N - Do not use the MPI datatype.
|
| 203 |
*/
|
| 204 |
if( ierr == MPI_SUCCESS )
|
| 205 |
ierr = MPI_Irecv( Mptr( U, 0, ibufR, LDU ), lengthR*LDU,
|
| 206 |
MPI_DOUBLE, partner, Cmsgid, comm, &request ); |
| 207 |
#endif
|
| 208 |
} |
| 209 |
|
| 210 |
if( lengthS > 0 ) |
| 211 |
{
|
| 212 |
#if 0
|
| 213 |
if( ierr == MPI_SUCCESS )
|
| 214 |
{
|
| 215 |
if( LDU == N )
|
| 216 |
ierr = MPI_Type_contiguous( lengthS*LDU, MPI_DOUBLE,
|
| 217 |
&type[I_SEND] );
|
| 218 |
else
|
| 219 |
ierr = MPI_Type_vector( lengthS, N, LDU, MPI_DOUBLE,
|
| 220 |
&type[I_SEND] );
|
| 221 |
}
|
| 222 |
if( ierr == MPI_SUCCESS )
|
| 223 |
ierr = MPI_Type_commit( &type[I_SEND] );
|
| 224 |
if( ierr == MPI_SUCCESS )
|
| 225 |
ierr = MPI_Send( Mptr( U, 0, ibufS, LDU ), 1, type[I_SEND],
|
| 226 |
partner, Cmsgid, comm );
|
| 227 |
if( ierr == MPI_SUCCESS )
|
| 228 |
ierr = MPI_Type_free( &type[I_SEND] );
|
| 229 |
#else
|
| 230 |
/*
|
| 231 |
* In our case, LDU is N - Do not use the MPI datatype.
|
| 232 |
*/
|
| 233 |
if( ierr == MPI_SUCCESS )
|
| 234 |
ierr = MPI_Send( Mptr( U, 0, ibufS, LDU ), lengthS*LDU,
|
| 235 |
MPI_DOUBLE, partner, Cmsgid, comm ); |
| 236 |
#endif
|
| 237 |
} |
| 238 |
|
| 239 |
if( lengthR > 0 ) |
| 240 |
{
|
| 241 |
if( ierr == MPI_SUCCESS )
|
| 242 |
ierr = MPI_Wait( &request, &status ); |
| 243 |
#if 0
|
| 244 |
if( ierr == MPI_SUCCESS )
|
| 245 |
ierr = MPI_Type_free( &type[I_RECV] );
|
| 246 |
#endif
|
| 247 |
} |
| 248 |
/*
|
| 249 |
* Probe for column panel - forward it when available
|
| 250 |
*/
|
| 251 |
if( *IFLAG == HPL_KEEP_TESTING ) (void) HPL_bcast( PBCST, IFLAG ); |
| 252 |
} |
| 253 |
|
| 254 |
if( ierr != MPI_SUCCESS )
|
| 255 |
{ HPL_pabort( __LINE__, "HPL_rollT", "MPI call failed" ); }
|
| 256 |
/*
|
| 257 |
* End of HPL_rollT
|
| 258 |
*/
|
| 259 |
} |