root / src / pgesv / HPL_pdgesvK1.c @ 9
Historique | Voir | Annoter | Télécharger (7,94 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 | void HPL_pdgesvK1
|
| 54 | 1 | equemene | ( |
| 55 | 1 | equemene | HPL_T_grid * GRID, |
| 56 | 1 | equemene | HPL_T_palg * ALGO, |
| 57 | 1 | equemene | HPL_T_pmat * A |
| 58 | 1 | equemene | ) |
| 59 | 1 | equemene | #else
|
| 60 | 1 | equemene | void HPL_pdgesvK1
|
| 61 | 1 | equemene | ( GRID, ALGO, A ) |
| 62 | 1 | equemene | HPL_T_grid * GRID; |
| 63 | 1 | equemene | HPL_T_palg * ALGO; |
| 64 | 1 | equemene | HPL_T_pmat * A; |
| 65 | 1 | equemene | #endif
|
| 66 | 1 | equemene | {
|
| 67 | 1 | equemene | /*
|
| 68 | 1 | equemene | * Purpose
|
| 69 | 1 | equemene | * =======
|
| 70 | 1 | equemene | *
|
| 71 | 1 | equemene | * HPL_pdgesvK1 factors a N+1-by-N matrix using LU factorization with row
|
| 72 | 1 | equemene | * partial pivoting. The main algorithm is the "right looking" variant
|
| 73 | 1 | equemene | * with look-ahead. The lower triangular factor is left unpivoted and
|
| 74 | 1 | equemene | * the pivots are not returned. The right hand side is the N+1 column of
|
| 75 | 1 | equemene | * the coefficient matrix.
|
| 76 | 1 | equemene | *
|
| 77 | 1 | equemene | * Arguments
|
| 78 | 1 | equemene | * =========
|
| 79 | 1 | equemene | *
|
| 80 | 1 | equemene | * GRID (local input) HPL_T_grid *
|
| 81 | 1 | equemene | * On entry, GRID points to the data structure containing the
|
| 82 | 1 | equemene | * process grid information.
|
| 83 | 1 | equemene | *
|
| 84 | 1 | equemene | * ALGO (global input) HPL_T_palg *
|
| 85 | 1 | equemene | * On entry, ALGO points to the data structure containing the
|
| 86 | 1 | equemene | * algorithmic parameters.
|
| 87 | 1 | equemene | *
|
| 88 | 1 | equemene | * A (local input/output) HPL_T_pmat *
|
| 89 | 1 | equemene | * On entry, A points to the data structure containing the local
|
| 90 | 1 | equemene | * array information.
|
| 91 | 1 | equemene | *
|
| 92 | 1 | equemene | * ---------------------------------------------------------------------
|
| 93 | 1 | equemene | */
|
| 94 | 1 | equemene | /*
|
| 95 | 1 | equemene | * .. Local Variables ..
|
| 96 | 1 | equemene | */
|
| 97 | 1 | equemene | HPL_T_panel * * panel = NULL;
|
| 98 | 1 | equemene | HPL_T_UPD_FUN HPL_pdupdate; |
| 99 | 1 | equemene | int N, depth, icurcol=0, j, jb, jj=0, jstart, |
| 100 | 1 | equemene | k, mycol, n, nb, nn, npcol, nq, |
| 101 | 1 | equemene | tag=MSGID_BEGIN_FACT, test=HPL_KEEP_TESTING; |
| 102 | 1 | equemene | /* ..
|
| 103 | 1 | equemene | * .. Executable Statements ..
|
| 104 | 1 | equemene | */
|
| 105 | 1 | equemene | mycol = GRID->mycol; npcol = GRID->npcol; |
| 106 | 1 | equemene | depth = ALGO->depth; HPL_pdupdate = ALGO->upfun; |
| 107 | 1 | equemene | N = A->n; nb = A->nb; |
| 108 | 1 | equemene | |
| 109 | 1 | equemene | if( N <= 0 ) return; |
| 110 | 1 | equemene | /*
|
| 111 | 1 | equemene | * Allocate a panel list of length depth + 1 (depth >= 1)
|
| 112 | 1 | equemene | */
|
| 113 | 1 | equemene | panel = (HPL_T_panel **)malloc( (size_t)(depth+1)*sizeof( HPL_T_panel *) ); |
| 114 | 1 | equemene | if( panel == NULL ) |
| 115 | 1 | equemene | { HPL_pabort( __LINE__, "HPL_pdgesvK1", "Memory allocation failed" ); }
|
| 116 | 1 | equemene | /*
|
| 117 | 1 | equemene | * Create and initialize the first depth panels
|
| 118 | 1 | equemene | */
|
| 119 | 1 | equemene | nq = HPL_numroc( N+1, nb, nb, mycol, 0, npcol ); nn = N; jstart = 0; |
| 120 | 1 | equemene | |
| 121 | 1 | equemene | for( k = 0; k < depth; k++ ) |
| 122 | 1 | equemene | {
|
| 123 | 1 | equemene | jb = Mmin( nn, nb ); |
| 124 | 1 | equemene | HPL_pdpanel_new( GRID, ALGO, nn, nn+1, jb, A, jstart, jstart,
|
| 125 | 1 | equemene | tag, &panel[k] ); |
| 126 | 1 | equemene | nn -= jb; jstart += jb; |
| 127 | 1 | equemene | if( mycol == icurcol ) { jj += jb; nq -= jb; }
|
| 128 | 1 | equemene | icurcol = MModAdd1( icurcol, npcol ); |
| 129 | 1 | equemene | tag = MNxtMgid( tag, MSGID_BEGIN_FACT, MSGID_END_FACT ); |
| 130 | 1 | equemene | } |
| 131 | 1 | equemene | /*
|
| 132 | 1 | equemene | * Initialize the lookahead - Factor jstart columns: panel[0..depth-1]
|
| 133 | 1 | equemene | */
|
| 134 | 1 | equemene | for( k = 0, j = 0; k < depth; k++ ) |
| 135 | 1 | equemene | {
|
| 136 | 1 | equemene | jb = jstart - j; jb = Mmin( jb, nb ); j += jb; |
| 137 | 1 | equemene | /*
|
| 138 | 1 | equemene | * Factor and broadcast k-th panel - use long topology for those
|
| 139 | 1 | equemene | */
|
| 140 | 1 | equemene | HPL_pdfact( panel[k] ); |
| 141 | 1 | equemene | (void) HPL_binit( panel[k] );
|
| 142 | 1 | equemene | do
|
| 143 | 1 | equemene | { (void) HPL_bcast( panel[k], &test ); }
|
| 144 | 1 | equemene | while( test != HPL_SUCCESS );
|
| 145 | 1 | equemene | (void) HPL_bwait( panel[k] );
|
| 146 | 1 | equemene | /*
|
| 147 | 1 | equemene | * Partial update of the depth-1-k panels in front of me
|
| 148 | 1 | equemene | */
|
| 149 | 1 | equemene | if( k < depth - 1 ) |
| 150 | 1 | equemene | {
|
| 151 | 1 | equemene | nn = HPL_numrocI( jstart-j, j, nb, nb, mycol, 0, npcol );
|
| 152 | 1 | equemene | HPL_pdupdate( NULL, NULL, panel[k], nn ); |
| 153 | 1 | equemene | } |
| 154 | 1 | equemene | } |
| 155 | 1 | equemene | /*
|
| 156 | 1 | equemene | * Main loop over the remaining columns of A
|
| 157 | 1 | equemene | */
|
| 158 | 1 | equemene | for( j = jstart; j < N; j += nb )
|
| 159 | 1 | equemene | {
|
| 160 | 1 | equemene | n = N - j; jb = Mmin( n, nb ); |
| 161 | 1 | equemene | /*
|
| 162 | 1 | equemene | * Allocate current panel resources - Finish latest update - Factor and
|
| 163 | 1 | equemene | * broadcast current panel
|
| 164 | 1 | equemene | */
|
| 165 | 1 | equemene | HPL_pdpanel_new( GRID, ALGO, n, n+1, jb, A, j, j, tag, &panel[depth] );
|
| 166 | 1 | equemene | |
| 167 | 1 | equemene | if( mycol == icurcol )
|
| 168 | 1 | equemene | {
|
| 169 | 1 | equemene | nn = HPL_numrocI( jb, j, nb, nb, mycol, 0, npcol );
|
| 170 | 1 | equemene | for( k = 0; k < depth; k++ ) /* partial updates 0..depth-1 */ |
| 171 | 1 | equemene | HPL_pdupdate( NULL, NULL, panel[k], nn ); |
| 172 | 1 | equemene | HPL_pdfact( panel[depth] ); /* factor current panel */
|
| 173 | 1 | equemene | } |
| 174 | 1 | equemene | else { nn = 0; } |
| 175 | 1 | equemene | /* Finish the latest update and broadcast the current panel */
|
| 176 | 1 | equemene | (void) HPL_binit( panel[depth] );
|
| 177 | 1 | equemene | HPL_pdupdate( panel[depth], &test, panel[0], nq-nn );
|
| 178 | 1 | equemene | (void) HPL_bwait( panel[depth] );
|
| 179 | 1 | equemene | /*
|
| 180 | 1 | equemene | * Release latest panel resources - circular of the panel pointers
|
| 181 | 1 | equemene | * Go to the next process row and column - update the message ids for
|
| 182 | 1 | equemene | * broadcast
|
| 183 | 1 | equemene | */
|
| 184 | 1 | equemene | (void) HPL_pdpanel_disp( &panel[0] ); |
| 185 | 1 | equemene | for( k = 0; k < depth; k++ ) panel[k] = panel[k+1]; |
| 186 | 1 | equemene | |
| 187 | 1 | equemene | if( mycol == icurcol ) { jj += jb; nq -= jb; }
|
| 188 | 1 | equemene | icurcol = MModAdd1( icurcol, npcol ); |
| 189 | 1 | equemene | tag = MNxtMgid( tag, MSGID_BEGIN_FACT, MSGID_END_FACT ); |
| 190 | 1 | equemene | } |
| 191 | 1 | equemene | /*
|
| 192 | 1 | equemene | * Clean-up: Finish updates - release panels and panel list
|
| 193 | 1 | equemene | */
|
| 194 | 1 | equemene | nn = HPL_numrocI( 1, N, nb, nb, mycol, 0, npcol ); |
| 195 | 1 | equemene | for( k = 0; k < depth; k++ ) |
| 196 | 1 | equemene | {
|
| 197 | 1 | equemene | HPL_pdupdate( NULL, NULL, panel[k], nn ); |
| 198 | 1 | equemene | (void) HPL_pdpanel_disp( &panel[k] );
|
| 199 | 1 | equemene | } |
| 200 | 1 | equemene | |
| 201 | 1 | equemene | if( panel ) free( panel );
|
| 202 | 1 | equemene | /*
|
| 203 | 1 | equemene | * End of HPL_pdgesvK1
|
| 204 | 1 | equemene | */
|
| 205 | 1 | equemene | } |