root / src / pgesv / HPL_pdlaswp00T.c @ 9
Historique | Voir | Annoter | Télécharger (16,8 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_pdlaswp00T
|
| 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_pdlaswp00T
|
| 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_pdlaswp00T 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 | * Bi-directional exchange is used to perform the swap :: broadcast of
|
| 77 | 1 | equemene | * the row panel U at once, resulting in a lower number of messages than
|
| 78 | 1 | equemene | * usual as well as a lower communication volume. With P process rows and
|
| 79 | 1 | equemene | * assuming bi-directional links, the running time of this function can
|
| 80 | 1 | equemene | * be approximated by:
|
| 81 | 1 | equemene | *
|
| 82 | 1 | equemene | * log_2(P) * (lat + 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. Mono
|
| 87 | 1 | equemene | * directional links will double this communication cost.
|
| 88 | 1 | equemene | *
|
| 89 | 1 | equemene | * Arguments
|
| 90 | 1 | equemene | * =========
|
| 91 | 1 | equemene | *
|
| 92 | 1 | equemene | * PBCST (local input/output) HPL_T_panel *
|
| 93 | 1 | equemene | * On entry, PBCST points to the data structure containing the
|
| 94 | 1 | equemene | * panel (to be broadcast) information.
|
| 95 | 1 | equemene | *
|
| 96 | 1 | equemene | * IFLAG (local input/output) int *
|
| 97 | 1 | equemene | * On entry, IFLAG indicates whether or not the broadcast has
|
| 98 | 1 | equemene | * already been completed. If not, probing will occur, and the
|
| 99 | 1 | equemene | * outcome will be contained in IFLAG on exit.
|
| 100 | 1 | equemene | *
|
| 101 | 1 | equemene | * PANEL (local input/output) HPL_T_panel *
|
| 102 | 1 | equemene | * On entry, PANEL points to the data structure containing the
|
| 103 | 1 | equemene | * panel (to be broadcast and swapped) information.
|
| 104 | 1 | equemene | *
|
| 105 | 1 | equemene | * NN (local input) const int
|
| 106 | 1 | equemene | * On entry, NN specifies the local number of columns of the
|
| 107 | 1 | equemene | * trailing submatrix to be swapped and broadcast starting at
|
| 108 | 1 | equemene | * the current position. NN must be at least zero.
|
| 109 | 1 | equemene | *
|
| 110 | 1 | equemene | * ---------------------------------------------------------------------
|
| 111 | 1 | equemene | */
|
| 112 | 1 | equemene | /*
|
| 113 | 1 | equemene | * .. Local Variables ..
|
| 114 | 1 | equemene | */
|
| 115 | 1 | equemene | MPI_Comm comm; |
| 116 | 1 | equemene | HPL_T_grid * grid; |
| 117 | 1 | equemene | double * A, * U, * W;
|
| 118 | 1 | equemene | void * vptr = NULL; |
| 119 | 1 | equemene | int * ipID, * lindxA, * lindxAU, * llen,
|
| 120 | 1 | equemene | * llen_sv; |
| 121 | 1 | equemene | unsigned int ip2, ip2_=1, ipdist, ipow=1, mask=1, |
| 122 | 1 | equemene | mydist, mydis_; |
| 123 | 1 | equemene | int Cmsgid=MSGID_BEGIN_PFACT, Np2, align,
|
| 124 | 1 | equemene | hdim, i, icurrow, *iflag, ipA, ipW, *ipl, |
| 125 | 1 | equemene | iprow, jb, k, lda, ldW, myrow, n, nprow, |
| 126 | 1 | equemene | partner, root, size_, usize; |
| 127 | 1 | equemene | #define LDU n
|
| 128 | 1 | equemene | /* ..
|
| 129 | 1 | equemene | * .. Executable Statements ..
|
| 130 | 1 | equemene | */
|
| 131 | 1 | equemene | n = Mmin( NN, PANEL->n ); jb = PANEL->jb; |
| 132 | 1 | equemene | /*
|
| 133 | 1 | equemene | * Quick return if there is nothing to do
|
| 134 | 1 | equemene | */
|
| 135 | 1 | equemene | if( ( n <= 0 ) || ( jb <= 0 ) ) return; |
| 136 | 1 | equemene | |
| 137 | 1 | equemene | #ifdef HPL_DETAILED_TIMING
|
| 138 | 1 | equemene | HPL_ptimer( HPL_TIMING_LASWP ); |
| 139 | 1 | equemene | #endif
|
| 140 | 1 | equemene | /*
|
| 141 | 1 | equemene | * Retrieve parameters from the PANEL data structure
|
| 142 | 1 | equemene | */
|
| 143 | 1 | equemene | grid = PANEL->grid; nprow = grid->nprow; myrow = grid->myrow; |
| 144 | 1 | equemene | comm = grid->col_comm; ip2 = (unsigned int)grid->row_ip2; |
| 145 | 1 | equemene | hdim = grid->row_hdim; align = PANEL->algo->align; |
| 146 | 1 | equemene | A = PANEL->A; U = PANEL->U; iflag = PANEL->IWORK; |
| 147 | 1 | equemene | lda = PANEL->lda; icurrow = PANEL->prow; usize = jb * n; |
| 148 | 1 | equemene | ldW = n + 1;
|
| 149 | 1 | equemene | /*
|
| 150 | 1 | equemene | * Allocate space for temporary W (ldW * jb)
|
| 151 | 1 | equemene | */
|
| 152 | 1 | equemene | vptr = (void*)malloc( ( (size_t)(align) +
|
| 153 | 1 | equemene | ((size_t)(jb) * (size_t)(ldW))) * |
| 154 | 1 | equemene | sizeof(double) ); |
| 155 | 1 | equemene | if( vptr == NULL ) |
| 156 | 1 | equemene | { HPL_pabort( __LINE__, "HPL_pdlaswp00T", "Memory allocation failed" ); }
|
| 157 | 1 | equemene | |
| 158 | 1 | equemene | W = (double *)HPL_PTR( vptr, ((size_t)(align) * sizeof(double) ) ); |
| 159 | 1 | equemene | /*
|
| 160 | 1 | equemene | * Construct ipID and its local counter parts lindxA, lindxAU - llen is
|
| 161 | 1 | equemene | * the number of rows/columns that I have in workspace and that I should
|
| 162 | 1 | equemene | * send. Compute lindx_, ipA, llen if it has not already been done for
|
| 163 | 1 | equemene | * this panel;
|
| 164 | 1 | equemene | */
|
| 165 | 1 | equemene | k = (int)((unsigned int)(jb) << 1); ipl = iflag + 1; ipID = ipl + 1; |
| 166 | 1 | equemene | lindxA = ipID + ((unsigned int)(k) << 1); lindxAU = lindxA + k; |
| 167 | 1 | equemene | llen = lindxAU + k; llen_sv = llen + nprow; |
| 168 | 1 | equemene | |
| 169 | 1 | equemene | if( *iflag == -1 ) /* no index arrays have been computed so far */ |
| 170 | 1 | equemene | {
|
| 171 | 1 | equemene | HPL_pipid( PANEL, ipl, ipID ); |
| 172 | 1 | equemene | HPL_plindx0( PANEL, *ipl, ipID, lindxA, lindxAU, llen_sv ); |
| 173 | 1 | equemene | *iflag = 0;
|
| 174 | 1 | equemene | } |
| 175 | 1 | equemene | else if( *iflag == 1 ) /* HPL_pdlaswp01T called before: reuse ipID */ |
| 176 | 1 | equemene | {
|
| 177 | 1 | equemene | HPL_plindx0( PANEL, *ipl, ipID, lindxA, lindxAU, llen_sv ); |
| 178 | 1 | equemene | *iflag = 0;
|
| 179 | 1 | equemene | } |
| 180 | 1 | equemene | /*
|
| 181 | 1 | equemene | * Copy the llen_sv into llen - Reset ipA to its correct value
|
| 182 | 1 | equemene | */
|
| 183 | 1 | equemene | ipA = llen_sv[myrow]; |
| 184 | 1 | equemene | for( i = 0; i < nprow; i++ ) { llen[i] = llen_sv[i]; } |
| 185 | 1 | equemene | /*
|
| 186 | 1 | equemene | * For i in [0..2*jb), lindxA[i] is the offset in A of a row that ulti-
|
| 187 | 1 | equemene | * mately goes to U( lindxAU[i], : ) or U( :, lindxAU[i] ). In icurrow,
|
| 188 | 1 | equemene | * we directly pack into U, otherwise we pack into workspace. The first
|
| 189 | 1 | equemene | * entry of each column packed in workspace is in fact the row or column
|
| 190 | 1 | equemene | * offset in U where it should go to.
|
| 191 | 1 | equemene | */
|
| 192 | 1 | equemene | if( myrow == icurrow )
|
| 193 | 1 | equemene | {
|
| 194 | 1 | equemene | HPL_dlaswp01T( ipA, n, A, lda, U, LDU, lindxA, lindxAU ); |
| 195 | 1 | equemene | } |
| 196 | 1 | equemene | else
|
| 197 | 1 | equemene | {
|
| 198 | 1 | equemene | HPL_dlaswp02N( ipA, n, A, lda, W, W+1, ldW, lindxA, lindxAU );
|
| 199 | 1 | equemene | } |
| 200 | 1 | equemene | /*
|
| 201 | 1 | equemene | * Probe for column panel - forward it when available
|
| 202 | 1 | equemene | */
|
| 203 | 1 | equemene | if( *IFLAG == HPL_KEEP_TESTING ) (void) HPL_bcast( PBCST, IFLAG ); |
| 204 | 1 | equemene | /*
|
| 205 | 1 | equemene | * Algorithm for bi-directional data exchange:
|
| 206 | 1 | equemene | *
|
| 207 | 1 | equemene | * As long as I have not talked to a process that already had the data
|
| 208 | 1 | equemene | * from icurrow, I will be sending the workspace, otherwise I will be
|
| 209 | 1 | equemene | * sending U. Note that the columns in workspace contain the local index
|
| 210 | 1 | equemene | * in U they should go to.
|
| 211 | 1 | equemene | *
|
| 212 | 1 | equemene | * If I am receiving from a process that has the data from icurrow, I
|
| 213 | 1 | equemene | * will be receiving in U, copy the data of U that stays into A, and
|
| 214 | 1 | equemene | * then the columns I have in workspace into U; otherwise I will be re-
|
| 215 | 1 | equemene | * ceiving in the remaining workspace. If I am one of those processes
|
| 216 | 1 | equemene | * that already has the data from icurrow, I will be immediately copying
|
| 217 | 1 | equemene | * the data I have in my workspace into U.
|
| 218 | 1 | equemene | *
|
| 219 | 1 | equemene | * When I receive U, some of U should be copied in my piece of A before
|
| 220 | 1 | equemene | * I can copy the rows I have in my workspace into U. This information
|
| 221 | 1 | equemene | * is kept in the lists lindx_: the row lindxAU[i] should be copied in
|
| 222 | 1 | equemene | * the row lindxA[i] of my piece of A, just as in the reversed initial
|
| 223 | 1 | equemene | * packing operation. Those rows are thus the first ones in the work ar-
|
| 224 | 1 | equemene | * ray. After this operation has been performed, I will not need
|
| 225 | 1 | equemene | * those lindx arrays, and I will always be sending a buffer of size
|
| 226 | 1 | equemene | * jb x n, or n x jb, that is, U.
|
| 227 | 1 | equemene | *
|
| 228 | 1 | equemene | * At every step of the algorithm, it is necesary to update the list
|
| 229 | 1 | equemene | * llen, so that I can figure out how large the next messages I will be
|
| 230 | 1 | equemene | * sending/receiving are. It is obvious when I am sending U. It is not
|
| 231 | 1 | equemene | * otherwise.
|
| 232 | 1 | equemene | *
|
| 233 | 1 | equemene | * We choose icurrow to be the source of the bi-directional exchange.
|
| 234 | 1 | equemene | * This allows the processes in the non-power 2 part to receive U at the
|
| 235 | 1 | equemene | * first exchange, and then broadcast internally this U so that those
|
| 236 | 1 | equemene | * processes can grab their piece of A.
|
| 237 | 1 | equemene | */
|
| 238 | 1 | equemene | if( myrow == icurrow ) { llen[myrow] = 0; ipA = 0; } |
| 239 | 1 | equemene | ipW = ipA; |
| 240 | 1 | equemene | Np2 = ( ( size_ = nprow - ip2 ) != 0 );
|
| 241 | 1 | equemene | mydist = (unsigned int)MModSub( myrow, icurrow, nprow ); |
| 242 | 1 | equemene | /*
|
| 243 | 1 | equemene | * bi-directional exchange: If nprow is not a power of 2, proc[i-ip2]
|
| 244 | 1 | equemene | * receives local data from proc[i] for all i in [ip2..nprow); icurrow
|
| 245 | 1 | equemene | * is the source, these last process indexes are relative to icurrow.
|
| 246 | 1 | equemene | */
|
| 247 | 1 | equemene | if( ( Np2 != 0 ) && ( ( partner = (int)(mydist ^ ip2) ) < nprow ) ) |
| 248 | 1 | equemene | {
|
| 249 | 1 | equemene | partner = MModAdd( icurrow, partner, nprow ); |
| 250 | 1 | equemene | |
| 251 | 1 | equemene | if( mydist == 0 ) /* I am the current row: I send U and recv W */ |
| 252 | 1 | equemene | {
|
| 253 | 1 | equemene | (void) HPL_sdrv( U, usize, Cmsgid, W, llen[partner] * ldW,
|
| 254 | 1 | equemene | Cmsgid, partner, comm ); |
| 255 | 1 | equemene | if( llen[partner] > 0 ) |
| 256 | 1 | equemene | HPL_dlaswp03T( llen[partner], n, U, LDU, W, W+1, ldW );
|
| 257 | 1 | equemene | } |
| 258 | 1 | equemene | else if( mydist == ip2 ) |
| 259 | 1 | equemene | { /* I recv U for later Bcast, I send my W */
|
| 260 | 1 | equemene | (void) HPL_sdrv( W, llen[myrow]*ldW, Cmsgid, U, usize,
|
| 261 | 1 | equemene | Cmsgid, partner, comm ); |
| 262 | 1 | equemene | } |
| 263 | 1 | equemene | else /* None of us is icurrow, we exchange our Ws */ |
| 264 | 1 | equemene | {
|
| 265 | 1 | equemene | if( ( mydist & ip2 ) != 0 ) |
| 266 | 1 | equemene | {
|
| 267 | 1 | equemene | (void) HPL_send( W, llen[myrow]*ldW, partner, Cmsgid, comm );
|
| 268 | 1 | equemene | } |
| 269 | 1 | equemene | else
|
| 270 | 1 | equemene | {
|
| 271 | 1 | equemene | (void) HPL_recv( Mptr( W, 0, ipW, ldW ), llen[partner]*ldW, |
| 272 | 1 | equemene | partner, Cmsgid, comm ); |
| 273 | 1 | equemene | if( llen[partner] > 0 ) ipW += llen[partner]; |
| 274 | 1 | equemene | } |
| 275 | 1 | equemene | } |
| 276 | 1 | equemene | } |
| 277 | 1 | equemene | /*
|
| 278 | 1 | equemene | * Update llen
|
| 279 | 1 | equemene | */
|
| 280 | 1 | equemene | for( i = 1; i < size_; i++ ) |
| 281 | 1 | equemene | {
|
| 282 | 1 | equemene | iprow = MModAdd( icurrow, i, nprow ); |
| 283 | 1 | equemene | partner = MModAdd( iprow, (int)(ip2), nprow );
|
| 284 | 1 | equemene | llen[ iprow ] += llen[ partner ]; |
| 285 | 1 | equemene | } |
| 286 | 1 | equemene | /*
|
| 287 | 1 | equemene | * Probe for column panel - forward it when available
|
| 288 | 1 | equemene | */
|
| 289 | 1 | equemene | if( *IFLAG == HPL_KEEP_TESTING ) (void) HPL_bcast( PBCST, IFLAG ); |
| 290 | 1 | equemene | /*
|
| 291 | 1 | equemene | * power of 2 part of the processes collection: only processes [0..ip2)
|
| 292 | 1 | equemene | * are working; some of them (mydist >> (k+1) == 0) either send or re-
|
| 293 | 1 | equemene | * ceive U. At every step k, k is in [0 .. hdim), of the algorithm, a
|
| 294 | 1 | equemene | * process pair that exchanges U is such that (mydist >> (k+1) == 0).
|
| 295 | 1 | equemene | * Among those processes, the ones that are sending U are such that
|
| 296 | 1 | equemene | * mydist >> k == 0.
|
| 297 | 1 | equemene | */
|
| 298 | 1 | equemene | if( mydist < ip2 )
|
| 299 | 1 | equemene | {
|
| 300 | 1 | equemene | k = 0;
|
| 301 | 1 | equemene | |
| 302 | 1 | equemene | while( k < hdim )
|
| 303 | 1 | equemene | {
|
| 304 | 1 | equemene | partner = (int)(mydist ^ ipow);
|
| 305 | 1 | equemene | partner = MModAdd( icurrow, partner, nprow ); |
| 306 | 1 | equemene | /*
|
| 307 | 1 | equemene | * Exchange and combine the local results - If I receive U, then I must
|
| 308 | 1 | equemene | * copy from U the rows that belong to my piece of A, and then update U
|
| 309 | 1 | equemene | * by copying in it the rows I have accumulated in W. Otherwise, I re-
|
| 310 | 1 | equemene | * ceive W. In this later case, and I have U, I shall update my copy of
|
| 311 | 1 | equemene | * U by copying in it the rows I have accumulated in W. If I did not
|
| 312 | 1 | equemene | * have U before, I simply need to update my pointer in W for later use.
|
| 313 | 1 | equemene | */
|
| 314 | 1 | equemene | if( ( mydist >> (unsigned int)( k + 1 ) ) == 0 ) |
| 315 | 1 | equemene | {
|
| 316 | 1 | equemene | if( ( mydist >> (unsigned int)(k) ) == 0 ) |
| 317 | 1 | equemene | {
|
| 318 | 1 | equemene | (void) HPL_sdrv( U, usize, Cmsgid, Mptr( W, 0, ipW, |
| 319 | 1 | equemene | ldW ), llen[partner]*ldW, Cmsgid, |
| 320 | 1 | equemene | partner, comm ); |
| 321 | 1 | equemene | HPL_dlaswp03T( llen[partner], n, U, LDU, Mptr( W, 0, ipW,
|
| 322 | 1 | equemene | ldW ), Mptr( W, 1, ipW, ldW ), ldW );
|
| 323 | 1 | equemene | ipW += llen[partner]; |
| 324 | 1 | equemene | } |
| 325 | 1 | equemene | else
|
| 326 | 1 | equemene | {
|
| 327 | 1 | equemene | (void) HPL_sdrv( W, llen[myrow]*ldW, Cmsgid, U, usize,
|
| 328 | 1 | equemene | Cmsgid, partner, comm ); |
| 329 | 1 | equemene | HPL_dlaswp04T( ipA, llen[myrow], n, U, LDU, A, lda, W, |
| 330 | 1 | equemene | W+1, ldW, lindxA, lindxAU );
|
| 331 | 1 | equemene | } |
| 332 | 1 | equemene | } |
| 333 | 1 | equemene | else
|
| 334 | 1 | equemene | {
|
| 335 | 1 | equemene | (void) HPL_sdrv( W, llen[myrow]*ldW, Cmsgid, Mptr( W, 0, |
| 336 | 1 | equemene | ipW, ldW ), llen[partner]*ldW, Cmsgid, |
| 337 | 1 | equemene | partner, comm ); |
| 338 | 1 | equemene | ipW += llen[partner]; |
| 339 | 1 | equemene | } |
| 340 | 1 | equemene | /*
|
| 341 | 1 | equemene | * Update llen - Go to next process pairs
|
| 342 | 1 | equemene | */
|
| 343 | 1 | equemene | iprow = icurrow; ipdist = 0;
|
| 344 | 1 | equemene | do
|
| 345 | 1 | equemene | {
|
| 346 | 1 | equemene | if( (unsigned int)( partner = (int)(ipdist ^ ipow) ) > ipdist ) |
| 347 | 1 | equemene | {
|
| 348 | 1 | equemene | partner = MModAdd( icurrow, partner, nprow ); |
| 349 | 1 | equemene | llen[iprow] += llen[partner]; |
| 350 | 1 | equemene | llen[partner] = llen[iprow]; |
| 351 | 1 | equemene | } |
| 352 | 1 | equemene | iprow = MModAdd( iprow, 1, nprow ); ipdist++;
|
| 353 | 1 | equemene | |
| 354 | 1 | equemene | } while( ipdist < ip2 );
|
| 355 | 1 | equemene | |
| 356 | 1 | equemene | ipow <<= 1; k++;
|
| 357 | 1 | equemene | /*
|
| 358 | 1 | equemene | * Probe for column panel - forward it when available
|
| 359 | 1 | equemene | */
|
| 360 | 1 | equemene | if( *IFLAG == HPL_KEEP_TESTING ) (void) HPL_bcast( PBCST, IFLAG ); |
| 361 | 1 | equemene | } |
| 362 | 1 | equemene | } |
| 363 | 1 | equemene | else
|
| 364 | 1 | equemene | {
|
| 365 | 1 | equemene | /*
|
| 366 | 1 | equemene | * non power of 2 part of the process collection: proc[ip2] broadcast U
|
| 367 | 1 | equemene | * to procs[ip2..nprow) (relatively to icurrow).
|
| 368 | 1 | equemene | */
|
| 369 | 1 | equemene | if( size_ > 1 ) |
| 370 | 1 | equemene | {
|
| 371 | 1 | equemene | k = size_ - 1;
|
| 372 | 1 | equemene | while( k > 1 ) { k >>= 1; ip2_ <<= 1; mask <<= 1; mask++; } |
| 373 | 1 | equemene | root = MModAdd( icurrow, (int)(ip2), nprow );
|
| 374 | 1 | equemene | mydis_ = (unsigned int)MModSub( myrow, root, nprow ); |
| 375 | 1 | equemene | |
| 376 | 1 | equemene | do
|
| 377 | 1 | equemene | {
|
| 378 | 1 | equemene | mask ^= ip2_; |
| 379 | 1 | equemene | if( ( mydis_ & mask ) == 0 ) |
| 380 | 1 | equemene | {
|
| 381 | 1 | equemene | partner = (int)(mydis_ ^ ip2_);
|
| 382 | 1 | equemene | if( ( mydis_ & ip2_ ) != 0 ) |
| 383 | 1 | equemene | {
|
| 384 | 1 | equemene | (void) HPL_recv( U, usize, MModAdd( root, partner,
|
| 385 | 1 | equemene | nprow ), Cmsgid, comm ); |
| 386 | 1 | equemene | |
| 387 | 1 | equemene | } |
| 388 | 1 | equemene | else if( partner < size_ ) |
| 389 | 1 | equemene | {
|
| 390 | 1 | equemene | (void) HPL_send( U, usize, MModAdd( root, partner,
|
| 391 | 1 | equemene | nprow ), Cmsgid, comm ); |
| 392 | 1 | equemene | } |
| 393 | 1 | equemene | } |
| 394 | 1 | equemene | ip2_ >>= 1;
|
| 395 | 1 | equemene | /*
|
| 396 | 1 | equemene | * Probe for column panel - forward it when available
|
| 397 | 1 | equemene | */
|
| 398 | 1 | equemene | if( *IFLAG == HPL_KEEP_TESTING ) (void) HPL_bcast( PBCST, IFLAG ); |
| 399 | 1 | equemene | |
| 400 | 1 | equemene | } while( ip2_ > 0 ); |
| 401 | 1 | equemene | } |
| 402 | 1 | equemene | /*
|
| 403 | 1 | equemene | * Every process in [ip2..nprow) (relatively to icurrow) grabs its piece
|
| 404 | 1 | equemene | * of A.
|
| 405 | 1 | equemene | */
|
| 406 | 1 | equemene | HPL_dlaswp05T( ipA, n, A, lda, U, LDU, lindxA, lindxAU ); |
| 407 | 1 | equemene | } |
| 408 | 1 | equemene | /*
|
| 409 | 1 | equemene | * If nprow is not a power of 2, proc[i-ip2] sends global result to
|
| 410 | 1 | equemene | * proc[i] for all i in [ip2..nprow);
|
| 411 | 1 | equemene | */
|
| 412 | 1 | equemene | if( ( Np2 != 0 ) && ( ( partner = (int)(mydist ^ ip2) ) < nprow ) ) |
| 413 | 1 | equemene | {
|
| 414 | 1 | equemene | partner = MModAdd( icurrow, partner, nprow ); |
| 415 | 1 | equemene | if( ( mydist & ip2 ) != 0 ) |
| 416 | 1 | equemene | { (void) HPL_recv( U, usize, partner, Cmsgid, comm ); }
|
| 417 | 1 | equemene | else
|
| 418 | 1 | equemene | { (void) HPL_send( U, usize, partner, Cmsgid, comm ); }
|
| 419 | 1 | equemene | } |
| 420 | 1 | equemene | |
| 421 | 1 | equemene | if( vptr ) free( vptr );
|
| 422 | 1 | equemene | /*
|
| 423 | 1 | equemene | * Probe for column panel - forward it when available
|
| 424 | 1 | equemene | */
|
| 425 | 1 | equemene | if( *IFLAG == HPL_KEEP_TESTING ) (void) HPL_bcast( PBCST, IFLAG ); |
| 426 | 1 | equemene | |
| 427 | 1 | equemene | #ifdef HPL_DETAILED_TIMING
|
| 428 | 1 | equemene | HPL_ptimer( HPL_TIMING_LASWP ); |
| 429 | 1 | equemene | #endif
|
| 430 | 1 | equemene | /*
|
| 431 | 1 | equemene | * End of HPL_pdlaswp00T
|
| 432 | 1 | equemene | */
|
| 433 | 1 | equemene | } |