root / src / pgesv / HPL_pdlaswp01N.c @ 8
Historique | Voir | Annoter | Télécharger (8,82 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_pdlaswp01N
|
| 54 | 1 | equemene | ( |
| 55 | 1 | equemene | HPL_T_panel * PBCST, |
| 56 | 1 | equemene | int * IFLAG,
|
| 57 | 1 | equemene | HPL_T_panel * PANEL, |
| 58 | 1 | equemene | const int NN |
| 59 | 1 | equemene | ) |
| 60 | 1 | equemene | #else
|
| 61 | 1 | equemene | void HPL_pdlaswp01N
|
| 62 | 1 | equemene | ( PBCST, IFLAG, PANEL, NN ) |
| 63 | 1 | equemene | HPL_T_panel * PBCST; |
| 64 | 1 | equemene | int * IFLAG;
|
| 65 | 1 | equemene | HPL_T_panel * PANEL; |
| 66 | 1 | equemene | const int NN; |
| 67 | 1 | equemene | #endif
|
| 68 | 1 | equemene | {
|
| 69 | 1 | equemene | /*
|
| 70 | 1 | equemene | * Purpose
|
| 71 | 1 | equemene | * =======
|
| 72 | 1 | equemene | *
|
| 73 | 1 | equemene | * HPL_pdlaswp01N applies the NB row interchanges to NN columns of the
|
| 74 | 1 | equemene | * trailing submatrix and broadcast a column panel.
|
| 75 | 1 | equemene | *
|
| 76 | 1 | equemene | * A "Spread then roll" algorithm performs the swap :: broadcast of the
|
| 77 | 1 | equemene | * row panel U at once, resulting in a minimal communication volume and
|
| 78 | 1 | equemene | * a "very good" use of the connectivity if available. With P process
|
| 79 | 1 | equemene | * rows and assuming bi-directional links, the running time of this
|
| 80 | 1 | equemene | * function can be approximated by:
|
| 81 | 1 | equemene | *
|
| 82 | 1 | equemene | * (log_2(P)+(P-1)) * lat + K * NB * LocQ(N) / bdwth
|
| 83 | 1 | equemene | *
|
| 84 | 1 | equemene | * where NB is the number of rows of the row panel U, N is the global
|
| 85 | 1 | equemene | * number of columns being updated, lat and bdwth are the latency and
|
| 86 | 1 | equemene | * bandwidth of the network for double precision real words. K is
|
| 87 | 1 | equemene | * a constant in (2,3] that depends on the achieved bandwidth during a
|
| 88 | 1 | equemene | * simultaneous message exchange between two processes. An empirical
|
| 89 | 1 | equemene | * optimistic value of K is typically 2.4.
|
| 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 information.
|
| 106 | 1 | equemene | *
|
| 107 | 1 | equemene | * NN (local input) const int
|
| 108 | 1 | equemene | * On entry, NN specifies the local number of columns of the
|
| 109 | 1 | equemene | * trailing submatrix to be swapped and broadcast starting at
|
| 110 | 1 | equemene | * the current position. NN must be at least zero.
|
| 111 | 1 | equemene | *
|
| 112 | 1 | equemene | * ---------------------------------------------------------------------
|
| 113 | 1 | equemene | */
|
| 114 | 1 | equemene | /*
|
| 115 | 1 | equemene | * .. Local Variables ..
|
| 116 | 1 | equemene | */
|
| 117 | 1 | equemene | double * A, * U;
|
| 118 | 1 | equemene | int * ipID, * iplen, * ipmap, * ipmapm1,
|
| 119 | 1 | equemene | * iwork, * lindxA = NULL, * lindxAU,
|
| 120 | 1 | equemene | * permU; |
| 121 | 1 | equemene | static int equil=-1; |
| 122 | 1 | equemene | int icurrow, * iflag, * ipA, * ipl, jb, k,
|
| 123 | 1 | equemene | lda, myrow, n, nprow; |
| 124 | 1 | equemene | #define LDU jb
|
| 125 | 1 | equemene | /* ..
|
| 126 | 1 | equemene | * .. Executable Statements ..
|
| 127 | 1 | equemene | */
|
| 128 | 1 | equemene | n = PANEL->n; n = Mmin( NN, n ); jb = PANEL->jb; |
| 129 | 1 | equemene | /*
|
| 130 | 1 | equemene | * Quick return if there is nothing to do
|
| 131 | 1 | equemene | */
|
| 132 | 1 | equemene | if( ( n <= 0 ) || ( jb <= 0 ) ) return; |
| 133 | 1 | equemene | #ifdef HPL_DETAILED_TIMING
|
| 134 | 1 | equemene | HPL_ptimer( HPL_TIMING_LASWP ); |
| 135 | 1 | equemene | #endif
|
| 136 | 1 | equemene | /*
|
| 137 | 1 | equemene | * Decide whether equilibration should be performed or not
|
| 138 | 1 | equemene | */
|
| 139 | 1 | equemene | if( equil == -1 ) equil = PANEL->algo->equil; |
| 140 | 1 | equemene | /*
|
| 141 | 1 | equemene | * Retrieve parameters from the PANEL data structure
|
| 142 | 1 | equemene | */
|
| 143 | 1 | equemene | nprow = PANEL->grid->nprow; myrow = PANEL->grid->myrow; |
| 144 | 1 | equemene | A = PANEL->A; U = PANEL->U; iflag = PANEL->IWORK; |
| 145 | 1 | equemene | lda = PANEL->lda; icurrow = PANEL->prow; |
| 146 | 1 | equemene | /*
|
| 147 | 1 | equemene | * Compute ipID (if not already done for this panel). lindxA and lindxAU
|
| 148 | 1 | equemene | * are of length at most 2*jb - iplen is of size nprow+1, ipmap, ipmapm1
|
| 149 | 1 | equemene | * are of size nprow, permU is of length jb, and this function needs a
|
| 150 | 1 | equemene | * workspace of size max( 2 * jb (plindx1), nprow+1(equil)):
|
| 151 | 1 | equemene | * 1(iflag) + 1(ipl) + 1(ipA) + 9*jb + 3*nprow + 1 + MAX(2*jb,nprow+1)
|
| 152 | 1 | equemene | * i.e. 4 + 9*jb + 3*nprow + max(2*jb, nprow+1);
|
| 153 | 1 | equemene | */
|
| 154 | 1 | equemene | k = (int)((unsigned int)(jb) << 1); ipl = iflag + 1; ipID = ipl + 1; |
| 155 | 1 | equemene | ipA = ipID + ((unsigned int)(k) << 1); lindxA = ipA + 1; |
| 156 | 1 | equemene | lindxAU = lindxA + k; iplen = lindxAU + k; ipmap = iplen + nprow + 1;
|
| 157 | 1 | equemene | ipmapm1 = ipmap + nprow; permU = ipmapm1 + nprow; iwork = permU + jb; |
| 158 | 1 | equemene | |
| 159 | 1 | equemene | if( *iflag == -1 ) /* no index arrays have been computed so far */ |
| 160 | 1 | equemene | {
|
| 161 | 1 | equemene | HPL_pipid( PANEL, ipl, ipID ); |
| 162 | 1 | equemene | HPL_plindx1( PANEL, *ipl, ipID, ipA, lindxA, lindxAU, iplen, |
| 163 | 1 | equemene | ipmap, ipmapm1, permU, iwork ); |
| 164 | 1 | equemene | *iflag = 1;
|
| 165 | 1 | equemene | } |
| 166 | 1 | equemene | else if( *iflag == 0 ) /* HPL_pdlaswp00N called before: reuse ipID */ |
| 167 | 1 | equemene | {
|
| 168 | 1 | equemene | HPL_plindx1( PANEL, *ipl, ipID, ipA, lindxA, lindxAU, iplen, |
| 169 | 1 | equemene | ipmap, ipmapm1, permU, iwork ); |
| 170 | 1 | equemene | *iflag = 1;
|
| 171 | 1 | equemene | } |
| 172 | 1 | equemene | else if( ( *iflag == 1 ) && ( equil != 0 ) ) |
| 173 | 1 | equemene | { /* HPL_pdlaswp01N was call before only re-compute IPLEN, IPMAP */
|
| 174 | 1 | equemene | HPL_plindx10( PANEL, *ipl, ipID, iplen, ipmap, ipmapm1 ); |
| 175 | 1 | equemene | *iflag = 1;
|
| 176 | 1 | equemene | } |
| 177 | 1 | equemene | /*
|
| 178 | 1 | equemene | * Copy into U the rows to be spread (local to icurrow)
|
| 179 | 1 | equemene | */
|
| 180 | 1 | equemene | if( myrow == icurrow )
|
| 181 | 1 | equemene | { HPL_dlaswp01N( *ipA, n, A, lda, U, LDU, lindxA, lindxAU ); }
|
| 182 | 1 | equemene | /*
|
| 183 | 1 | equemene | * Spread U - optionally probe for column panel
|
| 184 | 1 | equemene | */
|
| 185 | 1 | equemene | HPL_spreadN( PBCST, IFLAG, PANEL, HplRight, n, U, LDU, 0, iplen,
|
| 186 | 1 | equemene | ipmap, ipmapm1 ); |
| 187 | 1 | equemene | /*
|
| 188 | 1 | equemene | * Local exchange (everywhere but in process row icurrow)
|
| 189 | 1 | equemene | */
|
| 190 | 1 | equemene | if( myrow != icurrow )
|
| 191 | 1 | equemene | {
|
| 192 | 1 | equemene | k = ipmapm1[myrow]; |
| 193 | 1 | equemene | HPL_dlaswp06N( iplen[k+1]-iplen[k], n, A, lda, Mptr( U, iplen[k],
|
| 194 | 1 | equemene | 0, LDU ), LDU, lindxA );
|
| 195 | 1 | equemene | } |
| 196 | 1 | equemene | /*
|
| 197 | 1 | equemene | * Equilibration
|
| 198 | 1 | equemene | */
|
| 199 | 1 | equemene | if( equil != 0 ) |
| 200 | 1 | equemene | HPL_equil( PBCST, IFLAG, PANEL, HplNoTrans, n, U, LDU, iplen, |
| 201 | 1 | equemene | ipmap, ipmapm1, iwork ); |
| 202 | 1 | equemene | /*
|
| 203 | 1 | equemene | * Rolling phase
|
| 204 | 1 | equemene | */
|
| 205 | 1 | equemene | HPL_rollN( PBCST, IFLAG, PANEL, n, U, LDU, iplen, ipmap, ipmapm1 ); |
| 206 | 1 | equemene | /*
|
| 207 | 1 | equemene | * Permute U in every process row
|
| 208 | 1 | equemene | */
|
| 209 | 1 | equemene | HPL_dlaswp00N( jb, n, U, LDU, permU ); |
| 210 | 1 | equemene | |
| 211 | 1 | equemene | #ifdef HPL_DETAILED_TIMING
|
| 212 | 1 | equemene | HPL_ptimer( HPL_TIMING_LASWP ); |
| 213 | 1 | equemene | #endif
|
| 214 | 1 | equemene | /*
|
| 215 | 1 | equemene | * End of HPL_pdlaswp01N
|
| 216 | 1 | equemene | */
|
| 217 | 1 | equemene | } |