root / src / pgesv / HPL_equil.c @ 9
Historique | Voir | Annoter | Télécharger (10,38 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_equil
|
| 54 |
( |
| 55 |
HPL_T_panel * PBCST, |
| 56 |
int * IFLAG,
|
| 57 |
HPL_T_panel * PANEL, |
| 58 |
const enum HPL_TRANS TRANS, |
| 59 |
const int N, |
| 60 |
double * U,
|
| 61 |
const int LDU, |
| 62 |
int * IPLEN,
|
| 63 |
const int * IPMAP, |
| 64 |
const int * IPMAPM1, |
| 65 |
int * IWORK
|
| 66 |
) |
| 67 |
#else
|
| 68 |
void HPL_equil
|
| 69 |
( PBCST, IFLAG, PANEL, TRANS, N, U, LDU, IPLEN, IPMAP, IPMAPM1, IWORK ) |
| 70 |
HPL_T_panel * PBCST; |
| 71 |
int * IFLAG;
|
| 72 |
HPL_T_panel * PANEL; |
| 73 |
const enum HPL_TRANS TRANS; |
| 74 |
const int N; |
| 75 |
double * U;
|
| 76 |
const int LDU; |
| 77 |
int * IPLEN;
|
| 78 |
const int * IPMAP; |
| 79 |
const int * IPMAPM1; |
| 80 |
int * IWORK;
|
| 81 |
#endif
|
| 82 |
{
|
| 83 |
/*
|
| 84 |
* Purpose
|
| 85 |
* =======
|
| 86 |
*
|
| 87 |
* HPL_equil equilibrates the local pieces of U, so that on exit to
|
| 88 |
* this function, pieces of U contained in every process row are of the
|
| 89 |
* same size. This phase makes the rolling phase optimal. In addition,
|
| 90 |
* this function probes for the column panel L and forwards it when
|
| 91 |
* possible.
|
| 92 |
*
|
| 93 |
* Arguments
|
| 94 |
* =========
|
| 95 |
*
|
| 96 |
* PBCST (local input/output) HPL_T_panel *
|
| 97 |
* On entry, PBCST points to the data structure containing the
|
| 98 |
* panel (to be broadcast) information.
|
| 99 |
*
|
| 100 |
* IFLAG (local input/output) int *
|
| 101 |
* On entry, IFLAG indicates whether or not the broadcast has
|
| 102 |
* already been completed. If not, probing will occur, and the
|
| 103 |
* outcome will be contained in IFLAG on exit.
|
| 104 |
*
|
| 105 |
* PANEL (local input/output) HPL_T_panel *
|
| 106 |
* On entry, PANEL points to the data structure containing the
|
| 107 |
* panel (to be equilibrated) information.
|
| 108 |
*
|
| 109 |
* TRANS (global input) const enum HPL_TRANS
|
| 110 |
* On entry, TRANS specifies whether U is stored in transposed
|
| 111 |
* or non-transposed form.
|
| 112 |
*
|
| 113 |
* N (local input) const int
|
| 114 |
* On entry, N specifies the number of rows or columns of U. N
|
| 115 |
* must be at least 0.
|
| 116 |
*
|
| 117 |
* U (local input/output) double *
|
| 118 |
* On entry, U is an array of dimension (LDU,*) containing the
|
| 119 |
* local pieces of U in each process row.
|
| 120 |
*
|
| 121 |
* LDU (local input) const int
|
| 122 |
* On entry, LDU specifies the local leading dimension of U. LDU
|
| 123 |
* should be at least MAX(1,IPLEN[nprow]) when U is stored in
|
| 124 |
* non-transposed form, and MAX(1,N) otherwise.
|
| 125 |
*
|
| 126 |
* IPLEN (global input) int *
|
| 127 |
* On entry, IPLEN is an array of dimension NPROW+1. This array
|
| 128 |
* is such that IPLEN[i+1] - IPLEN[i] is the number of rows of U
|
| 129 |
* in process IPMAP[i].
|
| 130 |
*
|
| 131 |
* IPMAP (global input) const int *
|
| 132 |
* On entry, IPMAP is an array of dimension NPROW. This array
|
| 133 |
* contains the logarithmic mapping of the processes. In other
|
| 134 |
* words, IPMAP[myrow] is the absolute coordinate of the sorted
|
| 135 |
* process.
|
| 136 |
*
|
| 137 |
* IPMAPM1 (global input) const int *
|
| 138 |
* On entry, IPMAPM1 is an array of dimension NPROW. This array
|
| 139 |
* contains the inverse of the logarithmic mapping contained in
|
| 140 |
* IPMAP: For i in [0.. NPROCS) IPMAPM1[IPMAP[i]] = i.
|
| 141 |
*
|
| 142 |
* IWORK (workspace) int *
|
| 143 |
* On entry, IWORK is a workarray of dimension NPROW+1.
|
| 144 |
*
|
| 145 |
* ---------------------------------------------------------------------
|
| 146 |
*/
|
| 147 |
/*
|
| 148 |
* .. Local Variables ..
|
| 149 |
*/
|
| 150 |
int i, ip, ipU, ipcur, iprow, iptgt, lastrow,
|
| 151 |
left, npm1, nprow, ll, llU, llcur, lltgt, |
| 152 |
right, slen, smax, smin; |
| 153 |
/* ..
|
| 154 |
* .. Executable Statements ..
|
| 155 |
*/
|
| 156 |
if( ( npm1 = ( nprow = PANEL->grid->nprow ) - 1 ) <= 1 ) return; |
| 157 |
/*
|
| 158 |
* If the current distribution of the pieces of U is already optimal for
|
| 159 |
* the rolling phase, then return imediately. The optimal distribution
|
| 160 |
* is such that ip processes have smax items and the remaining processes
|
| 161 |
* only have smin items. Another way to check this is to verify that all
|
| 162 |
* differences IPLEN[i+1] - IPLEN[i] are either smin or smax.
|
| 163 |
*/
|
| 164 |
smax = ( ( slen = IPLEN[nprow] ) + npm1 ) / nprow; |
| 165 |
ip = slen - nprow * ( smin = slen / nprow ); |
| 166 |
|
| 167 |
iprow = 0;
|
| 168 |
do
|
| 169 |
{
|
| 170 |
ll = IPLEN[iprow+1] - IPLEN[iprow]; iprow++;
|
| 171 |
} while( ( iprow < nprow ) && ( ( ll == smin ) || ( ll == smax ) ) );
|
| 172 |
|
| 173 |
if( iprow == nprow ) return; |
| 174 |
/*
|
| 175 |
* Now, we are sure the distribution of the pieces of U is not optimal
|
| 176 |
* with respect to the rolling phase, thus perform equilibration. Go
|
| 177 |
* through the list of processes: Processes that have rows that do not
|
| 178 |
* belong to them with respect to the optimal mapping spread them in a
|
| 179 |
* logarithmic fashion. To simplify a little bit the implementation, and
|
| 180 |
* mainly the packing, a source process row spreads its data to its left
|
| 181 |
* first, and then to its right.
|
| 182 |
*/
|
| 183 |
IWORK[nprow] = slen; |
| 184 |
|
| 185 |
for( iprow = 0; iprow < nprow; iprow++ ) |
| 186 |
{
|
| 187 |
llU = IPLEN[iprow+1] - ( ipU = IPLEN[iprow] );
|
| 188 |
if( iprow < ip ) { lltgt = smax; iptgt = iprow * smax; }
|
| 189 |
else { lltgt = smin; iptgt = iprow * smin + ip; }
|
| 190 |
|
| 191 |
left = ( ipU < iptgt ); right = ( iptgt + lltgt < ipU + llU ); |
| 192 |
/*
|
| 193 |
* If I have something to spread to either the left or the right
|
| 194 |
*/
|
| 195 |
if( ( llU > 0 ) && ( left || right ) ) |
| 196 |
{ /* Figure out how much every other process should have */
|
| 197 |
|
| 198 |
ipcur = ipU; llcur = llU; |
| 199 |
|
| 200 |
for( i = 0; i < nprow; i++ ) |
| 201 |
{
|
| 202 |
if( i < ip ) { lltgt = smax; iptgt = i * smax; }
|
| 203 |
else { lltgt = smin; iptgt = i * smin + ip; }
|
| 204 |
lastrow = iptgt + lltgt - 1;
|
| 205 |
|
| 206 |
if( ( lastrow >= ipcur ) && ( llcur > 0 ) ) |
| 207 |
{ ll = lastrow - ipcur + 1; ll = Mmin( ll, llcur ); llcur -= ll; }
|
| 208 |
else { ll = 0; } |
| 209 |
|
| 210 |
IWORK[i] = ipcur; ipcur += ll; IWORK[i+1] = ipcur;
|
| 211 |
} |
| 212 |
/*
|
| 213 |
* Equilibration phase
|
| 214 |
*/
|
| 215 |
if( TRANS == HplNoTrans )
|
| 216 |
{
|
| 217 |
if( left )
|
| 218 |
{
|
| 219 |
HPL_spreadN( PBCST, IFLAG, PANEL, HplLeft, N, U, LDU, |
| 220 |
iprow, IWORK, IPMAP, IPMAPM1 ); |
| 221 |
} |
| 222 |
|
| 223 |
if( right )
|
| 224 |
{
|
| 225 |
HPL_spreadN( PBCST, IFLAG, PANEL, HplRight, N, U, LDU, |
| 226 |
iprow, IWORK, IPMAP, IPMAPM1 ); |
| 227 |
} |
| 228 |
} |
| 229 |
else
|
| 230 |
{
|
| 231 |
if( left )
|
| 232 |
{
|
| 233 |
HPL_spreadT( PBCST, IFLAG, PANEL, HplLeft, N, U, LDU, |
| 234 |
iprow, IWORK, IPMAP, IPMAPM1 ); |
| 235 |
} |
| 236 |
|
| 237 |
if( right )
|
| 238 |
{
|
| 239 |
HPL_spreadT( PBCST, IFLAG, PANEL, HplRight, N, U, LDU, |
| 240 |
iprow, IWORK, IPMAP, IPMAPM1 ); |
| 241 |
} |
| 242 |
} |
| 243 |
} |
| 244 |
} |
| 245 |
/*
|
| 246 |
* Finally update IPLEN with the indexes corresponding to the new dis-
|
| 247 |
* tribution of U - IPLEN[nprow] remained unchanged.
|
| 248 |
*/
|
| 249 |
for( i = 0; i < nprow; i++ ) IPLEN[i] = ( i < ip ? i*smax : i*smin + ip ); |
| 250 |
/*
|
| 251 |
* End of HPL_equil
|
| 252 |
*/
|
| 253 |
} |