root / src / pfact / HPL_pdmxswp.c @ 9
Historique | Voir | Annoter | Télécharger (12,31 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_pdmxswp
|
| 54 | 1 | equemene | ( |
| 55 | 1 | equemene | HPL_T_panel * PANEL, |
| 56 | 1 | equemene | const int M, |
| 57 | 1 | equemene | const int II, |
| 58 | 1 | equemene | const int JJ, |
| 59 | 1 | equemene | double * WORK
|
| 60 | 1 | equemene | ) |
| 61 | 1 | equemene | #else
|
| 62 | 1 | equemene | void HPL_pdmxswp
|
| 63 | 1 | equemene | ( PANEL, M, II, JJ, WORK ) |
| 64 | 1 | equemene | HPL_T_panel * PANEL; |
| 65 | 1 | equemene | const int M; |
| 66 | 1 | equemene | const int II; |
| 67 | 1 | equemene | const int JJ; |
| 68 | 1 | equemene | double * WORK;
|
| 69 | 1 | equemene | #endif
|
| 70 | 1 | equemene | {
|
| 71 | 1 | equemene | /*
|
| 72 | 1 | equemene | * Purpose
|
| 73 | 1 | equemene | * =======
|
| 74 | 1 | equemene | *
|
| 75 | 1 | equemene | * HPL_pdmxswp swaps and broadcasts the absolute value max row using
|
| 76 | 1 | equemene | * bi-directional exchange. The buffer is partially set by HPL_dlocmax.
|
| 77 | 1 | equemene | *
|
| 78 | 1 | equemene | * Bi-directional exchange is used to perform the swap::broadcast
|
| 79 | 1 | equemene | * operations at once for one column in the panel. This results in a
|
| 80 | 1 | equemene | * lower number of slightly larger messages than usual. On P processes
|
| 81 | 1 | equemene | * and assuming bi-directional links, the running time of this function
|
| 82 | 1 | equemene | * can be approximated by
|
| 83 | 1 | equemene | *
|
| 84 | 1 | equemene | * log_2( P ) * ( lat + ( 2 * N0 + 4 ) / bdwth )
|
| 85 | 1 | equemene | *
|
| 86 | 1 | equemene | * where lat and bdwth are the latency and bandwidth of the network for
|
| 87 | 1 | equemene | * double precision real elements. Communication only occurs in one
|
| 88 | 1 | equemene | * process column. Mono-directional links will cause the communication
|
| 89 | 1 | equemene | * cost to double.
|
| 90 | 1 | equemene | *
|
| 91 | 1 | equemene | * Arguments
|
| 92 | 1 | equemene | * =========
|
| 93 | 1 | equemene | *
|
| 94 | 1 | equemene | * PANEL (local input/output) HPL_T_panel *
|
| 95 | 1 | equemene | * On entry, PANEL points to the data structure containing the
|
| 96 | 1 | equemene | * panel information.
|
| 97 | 1 | equemene | *
|
| 98 | 1 | equemene | * M (local input) const int
|
| 99 | 1 | equemene | * On entry, M specifies the local number of rows of the matrix
|
| 100 | 1 | equemene | * column on which this function operates.
|
| 101 | 1 | equemene | *
|
| 102 | 1 | equemene | * II (local input) const int
|
| 103 | 1 | equemene | * On entry, II specifies the row offset where the column to be
|
| 104 | 1 | equemene | * operated on starts with respect to the panel.
|
| 105 | 1 | equemene | *
|
| 106 | 1 | equemene | * JJ (local input) const int
|
| 107 | 1 | equemene | * On entry, JJ specifies the column offset where the column to
|
| 108 | 1 | equemene | * be operated on starts with respect to the panel.
|
| 109 | 1 | equemene | *
|
| 110 | 1 | equemene | * WORK (local workspace) double *
|
| 111 | 1 | equemene | * On entry, WORK is a workarray of size at least 2 * (4+2*N0).
|
| 112 | 1 | equemene | * It is assumed that HPL_dlocmax was called prior to this
|
| 113 | 1 | equemene | * routine to initialize the first four entries of this array.
|
| 114 | 1 | equemene | * On exit, the N0 length max row is stored in WORK[4:4+N0-1];
|
| 115 | 1 | equemene | * Note that this is also the JJth row (or column) of L1. The
|
| 116 | 1 | equemene | * remaining part is used as a temporary array.
|
| 117 | 1 | equemene | *
|
| 118 | 1 | equemene | * ---------------------------------------------------------------------
|
| 119 | 1 | equemene | */
|
| 120 | 1 | equemene | /*
|
| 121 | 1 | equemene | * .. Local Variables ..
|
| 122 | 1 | equemene | */
|
| 123 | 1 | equemene | double gmax, tmp1;
|
| 124 | 1 | equemene | double * A0, * Wmx, * Wwork;
|
| 125 | 1 | equemene | HPL_T_grid * grid; |
| 126 | 1 | equemene | MPI_Comm comm; |
| 127 | 1 | equemene | unsigned int hdim, ip2, ip2_, ipow, k, mask; |
| 128 | 1 | equemene | int Np2, cnt_, cnt0, i, icurrow, lda, mydist,
|
| 129 | 1 | equemene | mydis_, myrow, n0, nprow, partner, rcnt, |
| 130 | 1 | equemene | root, scnt, size_; |
| 131 | 1 | equemene | /* ..
|
| 132 | 1 | equemene | * .. Executable Statements ..
|
| 133 | 1 | equemene | */
|
| 134 | 1 | equemene | #ifdef HPL_DETAILED_TIMING
|
| 135 | 1 | equemene | HPL_ptimer( HPL_TIMING_MXSWP ); |
| 136 | 1 | equemene | #endif
|
| 137 | 1 | equemene | grid = PANEL->grid; myrow = grid->myrow; nprow = grid->nprow; |
| 138 | 1 | equemene | /*
|
| 139 | 1 | equemene | * ip2 : the smallest power of two less than or equal to nprow;
|
| 140 | 1 | equemene | * hdim : dimension of the hypercube made of those ip2 processes;
|
| 141 | 1 | equemene | * Np2 : logical flag indicating whether or not nprow is a power of 2;
|
| 142 | 1 | equemene | */
|
| 143 | 1 | equemene | comm = grid->col_comm; ip2 = (unsigned int)(grid->row_ip2); |
| 144 | 1 | equemene | hdim = (unsigned int)(grid->row_hdim); n0 = PANEL->jb; |
| 145 | 1 | equemene | icurrow = PANEL->prow; Np2 = (int)( ( size_ = nprow - ip2 ) != 0 ); |
| 146 | 1 | equemene | mydist = MModSub( myrow, icurrow, nprow ); |
| 147 | 1 | equemene | /*
|
| 148 | 1 | equemene | * Set up pointers in workspace: WORK and Wwork point to the beginning
|
| 149 | 1 | equemene | * of the buffers of size 4 + 2*N0 to be combined. Wmx points to the row
|
| 150 | 1 | equemene | * owning the local (before combine) and global (after combine) absolute
|
| 151 | 1 | equemene | * value max. A0 points to the copy of the current row of the matrix.
|
| 152 | 1 | equemene | */
|
| 153 | 1 | equemene | cnt0 = ( cnt_ = n0 + 4 ) + n0; A0 = ( Wmx = WORK + 4 ) + n0; |
| 154 | 1 | equemene | Wwork = WORK + cnt0; |
| 155 | 1 | equemene | /*
|
| 156 | 1 | equemene | * Wmx[0:N0-1] := A[ilindx,0:N0-1] where ilindx is (int)(WORK[1]) (row
|
| 157 | 1 | equemene | * with max in current column). If I am the current process row, pack in
|
| 158 | 1 | equemene | * addition the current row of A in A0[0:N0-1]. If I do not own any row
|
| 159 | 1 | equemene | * of A, then zero out Wmx[0:N0-1].
|
| 160 | 1 | equemene | */
|
| 161 | 1 | equemene | if( M > 0 ) |
| 162 | 1 | equemene | {
|
| 163 | 1 | equemene | lda = PANEL->lda; |
| 164 | 1 | equemene | HPL_dcopy( n0, Mptr( PANEL->A, II+(int)(WORK[1]), 0, lda ), lda, |
| 165 | 1 | equemene | Wmx, 1 );
|
| 166 | 1 | equemene | if( myrow == icurrow )
|
| 167 | 1 | equemene | { HPL_dcopy( n0, Mptr( PANEL->A, II, 0, lda ), lda, A0, 1 ); }
|
| 168 | 1 | equemene | } |
| 169 | 1 | equemene | else { for( i = 0; i < n0; i++ ) Wmx[i] = HPL_rzero; } |
| 170 | 1 | equemene | /*
|
| 171 | 1 | equemene | * Combine the results (bi-directional exchange): the process coordina-
|
| 172 | 1 | equemene | * tes are relative to icurrow, this allows to reduce the communication
|
| 173 | 1 | equemene | * volume when nprow is not a power of 2.
|
| 174 | 1 | equemene | *
|
| 175 | 1 | equemene | * When nprow is not a power of 2: proc[i-ip2] receives local data from
|
| 176 | 1 | equemene | * proc[i] for all i in [ip2..nprow). In addition, proc[0] (icurrow)
|
| 177 | 1 | equemene | * sends to proc[ip2] the current row of A for later broadcast in procs
|
| 178 | 1 | equemene | * [ip2..nprow).
|
| 179 | 1 | equemene | */
|
| 180 | 1 | equemene | if( ( Np2 != 0 ) && |
| 181 | 1 | equemene | ( ( partner = (int)((unsigned int)(mydist) ^ ip2 ) ) < nprow ) ) |
| 182 | 1 | equemene | {
|
| 183 | 1 | equemene | if( ( mydist & ip2 ) != 0 ) |
| 184 | 1 | equemene | {
|
| 185 | 1 | equemene | if( mydist == (int)(ip2) ) |
| 186 | 1 | equemene | (void) HPL_sdrv( WORK, cnt_, MSGID_BEGIN_PFACT, A0, n0,
|
| 187 | 1 | equemene | MSGID_BEGIN_PFACT, MModAdd( partner, |
| 188 | 1 | equemene | icurrow, nprow ), comm ); |
| 189 | 1 | equemene | else
|
| 190 | 1 | equemene | (void) HPL_send( WORK, cnt_, MModAdd( partner, icurrow,
|
| 191 | 1 | equemene | nprow ), MSGID_BEGIN_PFACT, comm ); |
| 192 | 1 | equemene | } |
| 193 | 1 | equemene | else
|
| 194 | 1 | equemene | {
|
| 195 | 1 | equemene | if( mydist == 0 ) |
| 196 | 1 | equemene | (void) HPL_sdrv( A0, n0, MSGID_BEGIN_PFACT, Wwork, cnt_,
|
| 197 | 1 | equemene | MSGID_BEGIN_PFACT, MModAdd( partner, |
| 198 | 1 | equemene | icurrow, nprow ), comm ); |
| 199 | 1 | equemene | else
|
| 200 | 1 | equemene | (void) HPL_recv( Wwork, cnt_, MModAdd( partner, icurrow,
|
| 201 | 1 | equemene | nprow ), MSGID_BEGIN_PFACT, comm ); |
| 202 | 1 | equemene | |
| 203 | 1 | equemene | tmp1 = Mabs( Wwork[0] ); gmax = Mabs( WORK[0] ); |
| 204 | 1 | equemene | if( ( tmp1 > gmax ) ||
|
| 205 | 1 | equemene | ( ( tmp1 == gmax ) && ( Wwork[3] < WORK[3] ) ) ) |
| 206 | 1 | equemene | { HPL_dcopy( cnt_, Wwork, 1, WORK, 1 ); }
|
| 207 | 1 | equemene | } |
| 208 | 1 | equemene | } |
| 209 | 1 | equemene | |
| 210 | 1 | equemene | if( mydist < (int)(ip2) ) |
| 211 | 1 | equemene | {
|
| 212 | 1 | equemene | /*
|
| 213 | 1 | equemene | * power of 2 part of the processes collection: processes [0..ip2) are
|
| 214 | 1 | equemene | * combining (binary exchange); proc[0] has two rows to send, but one to
|
| 215 | 1 | equemene | * receive. At every step k in [0..hdim) of the algorithm, a process
|
| 216 | 1 | equemene | * pair exchanging 2 rows is such that myrow >> k+1 is 0. Among those
|
| 217 | 1 | equemene | * processes the ones that are sending one more row than what they are
|
| 218 | 1 | equemene | * receiving are such that myrow >> k is equal to 0.
|
| 219 | 1 | equemene | */
|
| 220 | 1 | equemene | k = 0; ipow = 1; |
| 221 | 1 | equemene | |
| 222 | 1 | equemene | while( k < hdim )
|
| 223 | 1 | equemene | {
|
| 224 | 1 | equemene | if( ( (unsigned int)(mydist) >> ( k + 1 ) ) == 0 ) |
| 225 | 1 | equemene | {
|
| 226 | 1 | equemene | if( ( (unsigned int)(mydist) >> k ) == 0 ) |
| 227 | 1 | equemene | { scnt = cnt0; rcnt = cnt_; }
|
| 228 | 1 | equemene | else
|
| 229 | 1 | equemene | { scnt = cnt_; rcnt = cnt0; }
|
| 230 | 1 | equemene | } |
| 231 | 1 | equemene | else { scnt = rcnt = cnt_; }
|
| 232 | 1 | equemene | |
| 233 | 1 | equemene | partner = (int)( (unsigned int)(mydist) ^ ipow ); |
| 234 | 1 | equemene | (void) HPL_sdrv( WORK, scnt, MSGID_BEGIN_PFACT, Wwork, rcnt,
|
| 235 | 1 | equemene | MSGID_BEGIN_PFACT, MModAdd( partner, icurrow, |
| 236 | 1 | equemene | nprow ), comm ); |
| 237 | 1 | equemene | |
| 238 | 1 | equemene | tmp1 = Mabs( Wwork[0] ); gmax = Mabs( WORK[0] ); |
| 239 | 1 | equemene | if( ( tmp1 > gmax ) ||
|
| 240 | 1 | equemene | ( ( tmp1 == gmax ) && ( Wwork[3] < WORK[3] ) ) ) |
| 241 | 1 | equemene | {
|
| 242 | 1 | equemene | HPL_dcopy( ( rcnt == cnt0 ? cnt0 : cnt_ ), Wwork, 1,
|
| 243 | 1 | equemene | WORK, 1 );
|
| 244 | 1 | equemene | } |
| 245 | 1 | equemene | else if( rcnt == cnt0 ) |
| 246 | 1 | equemene | { HPL_dcopy( n0, Wwork+cnt_, 1, A0, 1 ); }
|
| 247 | 1 | equemene | |
| 248 | 1 | equemene | ipow <<= 1; k++;
|
| 249 | 1 | equemene | } |
| 250 | 1 | equemene | } |
| 251 | 1 | equemene | else if( size_ > 1 ) |
| 252 | 1 | equemene | {
|
| 253 | 1 | equemene | /*
|
| 254 | 1 | equemene | * proc[ip2] broadcast current row of A to procs [ip2+1..nprow).
|
| 255 | 1 | equemene | */
|
| 256 | 1 | equemene | k = (unsigned int)(size_) - 1; ip2_ = mask = 1; |
| 257 | 1 | equemene | while( k > 1 ) { k >>= 1; ip2_ <<= 1; mask <<= 1; mask++; } |
| 258 | 1 | equemene | |
| 259 | 1 | equemene | root = MModAdd( icurrow, (int)(ip2), nprow );
|
| 260 | 1 | equemene | mydis_ = MModSub( myrow, root, nprow ); |
| 261 | 1 | equemene | |
| 262 | 1 | equemene | do
|
| 263 | 1 | equemene | {
|
| 264 | 1 | equemene | mask ^= ip2_; |
| 265 | 1 | equemene | if( ( mydis_ & mask ) == 0 ) |
| 266 | 1 | equemene | {
|
| 267 | 1 | equemene | partner = (int)(mydis_ ^ ip2_);
|
| 268 | 1 | equemene | if( ( mydis_ & ip2_ ) != 0 ) |
| 269 | 1 | equemene | {
|
| 270 | 1 | equemene | (void) HPL_recv( A0, n0, MModAdd( root, partner,
|
| 271 | 1 | equemene | nprow ), MSGID_BEGIN_PFACT, comm ); |
| 272 | 1 | equemene | } |
| 273 | 1 | equemene | else if( partner < size_ ) |
| 274 | 1 | equemene | {
|
| 275 | 1 | equemene | (void) HPL_send( A0, n0, MModAdd( root, partner,
|
| 276 | 1 | equemene | nprow ), MSGID_BEGIN_PFACT, comm ); |
| 277 | 1 | equemene | } |
| 278 | 1 | equemene | } |
| 279 | 1 | equemene | ip2_ >>= 1;
|
| 280 | 1 | equemene | } while( ip2_ > 0 ); |
| 281 | 1 | equemene | } |
| 282 | 1 | equemene | /*
|
| 283 | 1 | equemene | * If nprow is not a power of 2, for all i in [ip2..nprow), proc[i-ip2]
|
| 284 | 1 | equemene | * sends the pivot row to proc[i] along with the first four entries of
|
| 285 | 1 | equemene | * the WORK array.
|
| 286 | 1 | equemene | */
|
| 287 | 1 | equemene | if( ( Np2 != 0 ) && |
| 288 | 1 | equemene | ( ( partner = (int)((unsigned int)(mydist) ^ ip2 ) ) < nprow ) ) |
| 289 | 1 | equemene | {
|
| 290 | 1 | equemene | if( ( mydist & ip2 ) != 0 ) |
| 291 | 1 | equemene | {
|
| 292 | 1 | equemene | (void) HPL_recv( WORK, cnt_, MModAdd( partner, icurrow,
|
| 293 | 1 | equemene | nprow ), MSGID_BEGIN_PFACT, comm ); |
| 294 | 1 | equemene | } |
| 295 | 1 | equemene | else
|
| 296 | 1 | equemene | {
|
| 297 | 1 | equemene | (void) HPL_send( WORK, cnt_, MModAdd( partner, icurrow,
|
| 298 | 1 | equemene | nprow ), MSGID_BEGIN_PFACT, comm ); |
| 299 | 1 | equemene | } |
| 300 | 1 | equemene | } |
| 301 | 1 | equemene | /*
|
| 302 | 1 | equemene | * Save the global pivot index in pivot array
|
| 303 | 1 | equemene | */
|
| 304 | 1 | equemene | (PANEL->DPIV)[JJ] = WORK[2];
|
| 305 | 1 | equemene | #ifdef HPL_DETAILED_TIMING
|
| 306 | 1 | equemene | HPL_ptimer( HPL_TIMING_MXSWP ); |
| 307 | 1 | equemene | #endif
|
| 308 | 1 | equemene | /*
|
| 309 | 1 | equemene | * End of HPL_pdmxswp
|
| 310 | 1 | equemene | */
|
| 311 | 1 | equemene | } |