root / src / pgesv / HPL_plindx1.c
Historique | Voir | Annoter | Télécharger (11,39 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 |
#ifdef STDC_HEADERS
|
| 53 |
void HPL_plindx1
|
| 54 |
( |
| 55 |
HPL_T_panel * PANEL, |
| 56 |
const int K, |
| 57 |
const int * IPID, |
| 58 |
int * IPA,
|
| 59 |
int * LINDXA,
|
| 60 |
int * LINDXAU,
|
| 61 |
int * IPLEN,
|
| 62 |
int * IPMAP,
|
| 63 |
int * IPMAPM1,
|
| 64 |
int * PERMU,
|
| 65 |
int * IWORK
|
| 66 |
) |
| 67 |
#else
|
| 68 |
void HPL_plindx1
|
| 69 |
( PANEL, K, IPID, IPA, LINDXA, LINDXAU, IPLEN, IPMAP, IPMAPM1, PERMU, IWORK ) |
| 70 |
HPL_T_panel * PANEL; |
| 71 |
const int K; |
| 72 |
const int * IPID; |
| 73 |
int * IPA;
|
| 74 |
int * LINDXA;
|
| 75 |
int * LINDXAU;
|
| 76 |
int * IPLEN;
|
| 77 |
int * IPMAP;
|
| 78 |
int * IPMAPM1;
|
| 79 |
int * PERMU;
|
| 80 |
int * IWORK;
|
| 81 |
#endif
|
| 82 |
{
|
| 83 |
/*
|
| 84 |
* Purpose
|
| 85 |
* =======
|
| 86 |
*
|
| 87 |
* HPL_plindx1 computes two local arrays LINDXA and LINDXAU containing
|
| 88 |
* the local source and final destination position resulting from the
|
| 89 |
* application of row interchanges. In addition, this function computes
|
| 90 |
* three arrays IPLEN, IPMAP and IPMAPM1 that contain the logarithmic
|
| 91 |
* mapping information for the spreading phase.
|
| 92 |
*
|
| 93 |
* Arguments
|
| 94 |
* =========
|
| 95 |
*
|
| 96 |
* PANEL (local input/output) HPL_T_panel *
|
| 97 |
* On entry, PANEL points to the data structure containing the
|
| 98 |
* panel information.
|
| 99 |
*
|
| 100 |
* K (global input) const int
|
| 101 |
* On entry, K specifies the number of entries in IPID. K is at
|
| 102 |
* least 2*N, and at most 4*N.
|
| 103 |
*
|
| 104 |
* IPID (global input) const int *
|
| 105 |
* On entry, IPID is an array of length K. The first K entries
|
| 106 |
* of that array contain the src and final destination resulting
|
| 107 |
* from the application of the interchanges.
|
| 108 |
*
|
| 109 |
* IPA (global output) int *
|
| 110 |
* On exit, IPA specifies the number of rows that the current
|
| 111 |
* process row has that either belong to U or should be swapped
|
| 112 |
* with remote rows of A.
|
| 113 |
*
|
| 114 |
* LINDXA (global output) int *
|
| 115 |
* On entry, LINDXA is an array of dimension 2*N. On exit, this
|
| 116 |
* array contains the local indexes of the rows of A I have that
|
| 117 |
* should be copied into U.
|
| 118 |
*
|
| 119 |
* LINDXAU (global output) int *
|
| 120 |
* On exit, LINDXAU is an array of dimension 2*N. On exit, this
|
| 121 |
* array contains the local destination information encoded as
|
| 122 |
* follows. If LINDXAU(k) >= 0, row LINDXA(k) of A is to be
|
| 123 |
* copied in U at position LINDXAU(k). Otherwise, row LINDXA(k)
|
| 124 |
* of A should be locally copied into A(-LINDXAU(k),:).
|
| 125 |
*
|
| 126 |
* IPLEN (global output) int *
|
| 127 |
* On entry, IPLEN is an array of dimension NPROW + 1. On exit,
|
| 128 |
* this array is such that IPLEN[i] is the number of rows of A
|
| 129 |
* in the processes before process IPMAP[i] after the sort
|
| 130 |
* with the convention that IPLEN[nprow] is the total number of
|
| 131 |
* rows of the panel. In other words IPLEN[i+1]-IPLEN[i] is the
|
| 132 |
* local number of rows of A that should be moved to the process
|
| 133 |
* IPMAP[i]. IPLEN is such that the number of rows of the source
|
| 134 |
* process row can be computed as IPLEN[1] - IPLEN[0], and the
|
| 135 |
* remaining entries of this array are sorted so that the
|
| 136 |
* quantities IPLEN[i+1] - IPLEN[i] are logarithmically sorted.
|
| 137 |
*
|
| 138 |
* IPMAP (global output) int *
|
| 139 |
* On entry, IPMAP is an array of dimension NPROW. On exit, this
|
| 140 |
* array contains the logarithmic mapping of the processes. In
|
| 141 |
* other words, IPMAP[myrow] is the corresponding sorted process
|
| 142 |
* coordinate.
|
| 143 |
*
|
| 144 |
* IPMAPM1 (global output) int *
|
| 145 |
* On entry, IPMAPM1 is an array of dimension NPROW. On exit,
|
| 146 |
* this array contains the inverse of the logarithmic mapping
|
| 147 |
* contained in IPMAP: IPMAPM1[ IPMAP[i] ] = i, for all i in
|
| 148 |
* [0.. NPROCS)
|
| 149 |
*
|
| 150 |
* PERMU (global output) int *
|
| 151 |
* On entry, PERMU is an array of dimension JB. On exit, PERMU
|
| 152 |
* contains a sequence of permutations, that should be applied
|
| 153 |
* in increasing order to permute in place the row panel U.
|
| 154 |
*
|
| 155 |
* IWORK (workspace) int *
|
| 156 |
* On entry, IWORK is a workarray of dimension 2*JB.
|
| 157 |
*
|
| 158 |
* ---------------------------------------------------------------------
|
| 159 |
*/
|
| 160 |
/*
|
| 161 |
* .. Local Variables ..
|
| 162 |
*/
|
| 163 |
int * iwork;
|
| 164 |
int dst, dstrow, fndd, i, ia, icurrow, il,
|
| 165 |
ip, ipU, iroff, j, jb, myrow, nb, nprow, |
| 166 |
src, srcrow; |
| 167 |
/* ..
|
| 168 |
* .. Executable Statements ..
|
| 169 |
*/
|
| 170 |
/*
|
| 171 |
* Logarithmic sort of the processes - compute IPMAP, IPLEN and IPMAPM1
|
| 172 |
*/
|
| 173 |
HPL_plindx10( PANEL, K, IPID, IPLEN, IPMAP, IPMAPM1 ); |
| 174 |
/*
|
| 175 |
* Compute the local arrays LINDXA and LINDXAU containing the local
|
| 176 |
* source and final destination position resulting from the application
|
| 177 |
* of N interchanges. Compute LINDXA and LINDXAU in icurrow, and LINDXA
|
| 178 |
* elsewhere and PERMU in every process.
|
| 179 |
*/
|
| 180 |
myrow = PANEL->grid->myrow; nprow = PANEL->grid->nprow; |
| 181 |
jb = PANEL->jb; nb = PANEL->nb; ia = PANEL->ia; |
| 182 |
iroff = PANEL->ii; icurrow = PANEL->prow; |
| 183 |
|
| 184 |
iwork = IWORK + jb; |
| 185 |
|
| 186 |
if( myrow == icurrow )
|
| 187 |
{
|
| 188 |
for( i = 0, ip = 0, ipU = 0; i < K; i += 2 ) |
| 189 |
{
|
| 190 |
src = IPID[i]; Mindxg2p( src, nb, nb, srcrow, 0, nprow );
|
| 191 |
|
| 192 |
if( srcrow == icurrow )
|
| 193 |
{
|
| 194 |
dst = IPID[i+1]; Mindxg2p( dst, nb, nb, dstrow, 0, nprow ); |
| 195 |
|
| 196 |
Mindxg2l( il, src, nb, nb, myrow, 0, nprow );
|
| 197 |
LINDXA[ip] = il - iroff; |
| 198 |
|
| 199 |
if( ( dstrow == icurrow ) && ( dst - ia < jb ) )
|
| 200 |
{
|
| 201 |
PERMU[ipU] = dst - ia; il = IPMAPM1[dstrow]; |
| 202 |
j = IPLEN[il]; iwork[ipU] = LINDXAU[ip] = j; |
| 203 |
IPLEN[il]++; ipU++; |
| 204 |
} |
| 205 |
else if( dstrow != icurrow ) |
| 206 |
{
|
| 207 |
j = 0;
|
| 208 |
do { fndd = ( dst == IPID[j] ); j+=2; } |
| 209 |
while( !fndd && ( j < K ) );
|
| 210 |
|
| 211 |
PERMU[ipU] = IPID[j-1]-ia; il = IPMAPM1[dstrow];
|
| 212 |
j = IPLEN[il]; iwork[ipU] = LINDXAU[ip] = j; |
| 213 |
IPLEN[il]++; ipU++; |
| 214 |
} |
| 215 |
else if( ( dstrow == icurrow ) && ( dst - ia >= jb ) ) |
| 216 |
{
|
| 217 |
Mindxg2l( il, dst, nb, nb, myrow, 0, nprow );
|
| 218 |
LINDXAU[ip] = iroff - il; |
| 219 |
} |
| 220 |
ip++; |
| 221 |
} |
| 222 |
} |
| 223 |
*IPA = ip; |
| 224 |
} |
| 225 |
else
|
| 226 |
{
|
| 227 |
for( i = 0, ip = 0, ipU = 0; i < K; i += 2 ) |
| 228 |
{
|
| 229 |
src = IPID[i ]; Mindxg2p( src, nb, nb, srcrow, 0, nprow );
|
| 230 |
dst = IPID[i+1]; Mindxg2p( dst, nb, nb, dstrow, 0, nprow ); |
| 231 |
/*
|
| 232 |
* LINDXA[i] is the local index of the row of A that belongs into U
|
| 233 |
*/
|
| 234 |
if( myrow == dstrow )
|
| 235 |
{
|
| 236 |
Mindxg2l( il, dst, nb, nb, myrow, 0, nprow );
|
| 237 |
LINDXA[ip] = il - iroff; ip++; |
| 238 |
} |
| 239 |
/*
|
| 240 |
* iwork[i] is the local (current) position index in U
|
| 241 |
* PERMU[i] is the local (final) destination index in U
|
| 242 |
*/
|
| 243 |
if( srcrow == icurrow )
|
| 244 |
{
|
| 245 |
if( ( dstrow == icurrow ) && ( dst - ia < jb ) )
|
| 246 |
{
|
| 247 |
PERMU[ipU] = dst - ia; il = IPMAPM1[dstrow]; |
| 248 |
iwork[ipU] = IPLEN[il]; IPLEN[il]++; ipU++; |
| 249 |
} |
| 250 |
else if( dstrow != icurrow ) |
| 251 |
{
|
| 252 |
j = 0;
|
| 253 |
do { fndd = ( dst == IPID[j] ); j+=2; } |
| 254 |
while( !fndd && ( j < K ) );
|
| 255 |
PERMU[ipU] = IPID[j-1] - ia; il = IPMAPM1[dstrow];
|
| 256 |
iwork[ipU] = IPLEN[il]; IPLEN[il]++; ipU++; |
| 257 |
} |
| 258 |
} |
| 259 |
} |
| 260 |
*IPA = 0;
|
| 261 |
} |
| 262 |
/*
|
| 263 |
* Simplify iwork and PERMU, return in PERMU the sequence of permutation
|
| 264 |
* that need to be apply to U after it has been broadcast.
|
| 265 |
*/
|
| 266 |
HPL_perm( jb, iwork, PERMU, IWORK ); |
| 267 |
/*
|
| 268 |
* Reset IPLEN to its correct value
|
| 269 |
*/
|
| 270 |
for( i = nprow; i > 0; i-- ) IPLEN[i] = IPLEN[i-1]; |
| 271 |
IPLEN[0] = 0; |
| 272 |
/*
|
| 273 |
* End of HPL_plindx1
|
| 274 |
*/
|
| 275 |
} |