root / src / pgesv / HPL_spreadT.c @ 9
Historique | Voir | Annoter | Télécharger (14,74 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_spreadT
|
| 54 |
( |
| 55 |
HPL_T_panel * PBCST, |
| 56 |
int * IFLAG,
|
| 57 |
HPL_T_panel * PANEL, |
| 58 |
const enum HPL_SIDE SIDE, |
| 59 |
const int N, |
| 60 |
double * U,
|
| 61 |
const int LDU, |
| 62 |
const int SRCDIST, |
| 63 |
const int * IPLEN, |
| 64 |
const int * IPMAP, |
| 65 |
const int * IPMAPM1 |
| 66 |
) |
| 67 |
#else
|
| 68 |
void HPL_spreadT
|
| 69 |
( PBCST, IFLAG, PANEL, SIDE, N, U, LDU, SRCDIST, IPLEN, IPMAP, IPMAPM1 ) |
| 70 |
HPL_T_panel * PBCST; |
| 71 |
int * IFLAG;
|
| 72 |
HPL_T_panel * PANEL; |
| 73 |
const enum HPL_SIDE SIDE; |
| 74 |
const int N; |
| 75 |
double * U;
|
| 76 |
const int LDU; |
| 77 |
const int SRCDIST; |
| 78 |
const int * IPLEN; |
| 79 |
const int * IPMAP; |
| 80 |
const int * IPMAPM1; |
| 81 |
#endif
|
| 82 |
{
|
| 83 |
/*
|
| 84 |
* Purpose
|
| 85 |
* =======
|
| 86 |
*
|
| 87 |
* HPL_spreadT spreads the local array containing local pieces of U, so
|
| 88 |
* that on exit to this function, a piece of U is contained in every
|
| 89 |
* process row. The array IPLEN contains the number of columns of U,
|
| 90 |
* that should be spread on any given process row. This function also
|
| 91 |
* probes for the presence of the column panel PBCST. If available,
|
| 92 |
* this panel will be forwarded. If PBCST is NULL on input, this
|
| 93 |
* probing mechanism will be disabled.
|
| 94 |
*
|
| 95 |
* Arguments
|
| 96 |
* =========
|
| 97 |
*
|
| 98 |
* PBCST (local input/output) HPL_T_panel *
|
| 99 |
* On entry, PBCST points to the data structure containing the
|
| 100 |
* panel (to be broadcast) information.
|
| 101 |
*
|
| 102 |
* IFLAG (local input/output) int *
|
| 103 |
* On entry, IFLAG indicates whether or not the broadcast has
|
| 104 |
* already been completed. If not, probing will occur, and the
|
| 105 |
* outcome will be contained in IFLAG on exit.
|
| 106 |
*
|
| 107 |
* PANEL (local input/output) HPL_T_panel *
|
| 108 |
* On entry, PANEL points to the data structure containing the
|
| 109 |
* panel (to be spread) information.
|
| 110 |
*
|
| 111 |
* SIDE (global input) const enum HPL_SIDE
|
| 112 |
* On entry, SIDE specifies whether the local piece of U located
|
| 113 |
* in process IPMAP[SRCDIST] should be spread to the right or to
|
| 114 |
* the left. This feature is used by the equilibration process.
|
| 115 |
*
|
| 116 |
* N (global input) const int
|
| 117 |
* On entry, N specifies the local number of rows of U. N must
|
| 118 |
* be at least zero.
|
| 119 |
*
|
| 120 |
* U (local input/output) double *
|
| 121 |
* On entry, U is an array of dimension (LDU,*) containing the
|
| 122 |
* local pieces of U.
|
| 123 |
*
|
| 124 |
* LDU (local input) const int
|
| 125 |
* On entry, LDU specifies the local leading dimension of U. LDU
|
| 126 |
* should be at least MAX(1,N).
|
| 127 |
*
|
| 128 |
* SRCDIST (local input) const int
|
| 129 |
* On entry, SRCDIST specifies the source process that spreads
|
| 130 |
* its piece of U.
|
| 131 |
*
|
| 132 |
* IPLEN (global input) const int *
|
| 133 |
* On entry, IPLEN is an array of dimension NPROW+1. This array
|
| 134 |
* is such that IPLEN[i+1] - IPLEN[i] is the number of rows of U
|
| 135 |
* in each process before process IPMAP[i], with the convention
|
| 136 |
* that IPLEN[nprow] is the total number of rows. In other words
|
| 137 |
* IPLEN[i+1] - IPLEN[i] is the local number of rows of U that
|
| 138 |
* should be moved to process IPMAP[i].
|
| 139 |
*
|
| 140 |
* IPMAP (global input) const int *
|
| 141 |
* On entry, IPMAP is an array of dimension NPROW. This array
|
| 142 |
* contains the logarithmic mapping of the processes. In other
|
| 143 |
* words, IPMAP[myrow] is the absolute coordinate of the sorted
|
| 144 |
* process.
|
| 145 |
*
|
| 146 |
* IPMAPM1 (global input) const int *
|
| 147 |
* On entry, IPMAPM1 is an array of dimension NPROW. This array
|
| 148 |
* contains the inverse of the logarithmic mapping contained in
|
| 149 |
* IPMAP: For i in [0.. NPROW) IPMAPM1[IPMAP[i]] = i.
|
| 150 |
*
|
| 151 |
* ---------------------------------------------------------------------
|
| 152 |
*/
|
| 153 |
/*
|
| 154 |
* .. Local Variables ..
|
| 155 |
*/
|
| 156 |
#if 0
|
| 157 |
MPI_Datatype type;
|
| 158 |
#endif
|
| 159 |
MPI_Status status; |
| 160 |
MPI_Comm comm; |
| 161 |
unsigned int ip2=1, mask=1, mydist, mydist2; |
| 162 |
int Cmsgid=MSGID_BEGIN_PFACT, ibuf,
|
| 163 |
ierr=MPI_SUCCESS, il, k, lbuf, lgth, myrow, |
| 164 |
npm1, nprow, partner; |
| 165 |
/* ..
|
| 166 |
* .. Executable Statements ..
|
| 167 |
*/
|
| 168 |
myrow = PANEL->grid->myrow; nprow = PANEL->grid->nprow; |
| 169 |
comm = PANEL->grid->col_comm; |
| 170 |
/*
|
| 171 |
* Spread U
|
| 172 |
*/
|
| 173 |
if( SIDE == HplLeft )
|
| 174 |
{
|
| 175 |
nprow = ( npm1 = SRCDIST ) + 1;
|
| 176 |
if( ( ( mydist = (unsigned int)(IPMAPM1[myrow]) ) > |
| 177 |
(unsigned int)(SRCDIST) ) || ( npm1 == 0 ) ) return; |
| 178 |
|
| 179 |
k = npm1; while( k > 1 ) { k >>= 1; ip2 <<= 1; mask <<= 1; mask++; } |
| 180 |
mydist2 = ( mydist = npm1 - mydist ); il = npm1 - ip2; |
| 181 |
lgth = IPLEN[nprow]; |
| 182 |
|
| 183 |
do
|
| 184 |
{
|
| 185 |
mask ^= ip2; |
| 186 |
|
| 187 |
if( ( mydist & mask ) == 0 ) |
| 188 |
{
|
| 189 |
lbuf = IPLEN[il+1] - ( ibuf = IPLEN[il-Mmin(il, (int)(ip2))] ); |
| 190 |
|
| 191 |
if( lbuf > 0 ) |
| 192 |
{
|
| 193 |
partner = mydist ^ ip2; |
| 194 |
|
| 195 |
if( mydist & ip2 )
|
| 196 |
{
|
| 197 |
#if 0
|
| 198 |
if( ierr == MPI_SUCCESS )
|
| 199 |
{
|
| 200 |
if( LDU == N )
|
| 201 |
ierr = MPI_Type_contiguous( lbuf*LDU, MPI_DOUBLE,
|
| 202 |
&type );
|
| 203 |
else
|
| 204 |
ierr = MPI_Type_vector( lbuf, N, LDU, MPI_DOUBLE,
|
| 205 |
&type );
|
| 206 |
}
|
| 207 |
if( ierr == MPI_SUCCESS )
|
| 208 |
ierr = MPI_Type_commit( &type );
|
| 209 |
if( ierr == MPI_SUCCESS )
|
| 210 |
ierr = MPI_Recv( Mptr( U, 0, ibuf, LDU ), 1, type,
|
| 211 |
IPMAP[npm1-partner], Cmsgid, comm,
|
| 212 |
&status );
|
| 213 |
if( ierr == MPI_SUCCESS )
|
| 214 |
ierr = MPI_Type_free( &type );
|
| 215 |
#else
|
| 216 |
/*
|
| 217 |
* In our case, LDU is N - do not use the MPI Datatypes
|
| 218 |
*/
|
| 219 |
if( ierr == MPI_SUCCESS )
|
| 220 |
ierr = MPI_Recv( Mptr( U, 0, ibuf, LDU ), lbuf*N,
|
| 221 |
MPI_DOUBLE, IPMAP[npm1-partner], |
| 222 |
Cmsgid, comm, &status ); |
| 223 |
#endif
|
| 224 |
} |
| 225 |
else if( partner < nprow ) |
| 226 |
{
|
| 227 |
#if 0
|
| 228 |
if( ierr == MPI_SUCCESS )
|
| 229 |
{
|
| 230 |
if( LDU == N )
|
| 231 |
ierr = MPI_Type_contiguous( lbuf*LDU, MPI_DOUBLE,
|
| 232 |
&type );
|
| 233 |
else
|
| 234 |
ierr = MPI_Type_vector( lbuf, N, LDU, MPI_DOUBLE,
|
| 235 |
&type );
|
| 236 |
}
|
| 237 |
if( ierr == MPI_SUCCESS )
|
| 238 |
ierr = MPI_Type_commit( &type );
|
| 239 |
if( ierr == MPI_SUCCESS )
|
| 240 |
ierr = MPI_Send( Mptr( U, 0, ibuf, LDU ), 1, type,
|
| 241 |
IPMAP[npm1-partner], Cmsgid, comm );
|
| 242 |
if( ierr == MPI_SUCCESS )
|
| 243 |
ierr = MPI_Type_free( &type );
|
| 244 |
#else
|
| 245 |
/*
|
| 246 |
* In our case, LDU is N - do not use the MPI Datatypes
|
| 247 |
*/
|
| 248 |
if( ierr == MPI_SUCCESS )
|
| 249 |
ierr = MPI_Send( Mptr( U, 0, ibuf, LDU ), lbuf*N,
|
| 250 |
MPI_DOUBLE, IPMAP[npm1-partner], |
| 251 |
Cmsgid, comm ); |
| 252 |
#endif
|
| 253 |
} |
| 254 |
} |
| 255 |
} |
| 256 |
|
| 257 |
if( mydist2 < ip2 ) { ip2 >>= 1; il += ip2; } |
| 258 |
else { mydist2 -= ip2; ip2 >>= 1; il -= ip2; } |
| 259 |
/*
|
| 260 |
* Probe for column panel - forward it when available
|
| 261 |
*/
|
| 262 |
if( *IFLAG == HPL_KEEP_TESTING ) (void) HPL_bcast( PBCST, IFLAG ); |
| 263 |
|
| 264 |
} while( ip2 > 0 ); |
| 265 |
} |
| 266 |
else
|
| 267 |
{
|
| 268 |
npm1 = ( nprow -= SRCDIST ) - 1;
|
| 269 |
if( ( ( mydist = (unsigned int)(IPMAPM1[myrow]) ) < |
| 270 |
(unsigned int)(SRCDIST) ) || ( npm1 == 0 ) ) return; |
| 271 |
|
| 272 |
k = npm1; while( k > 1 ) { k >>= 1; ip2 <<= 1; mask <<= 1; mask++; } |
| 273 |
mydist2 = ( mydist -= SRCDIST ); il = ip2; |
| 274 |
/*
|
| 275 |
* Spread to the right - offset the IPLEN and IPMAP arrays
|
| 276 |
*/
|
| 277 |
lgth = IPLEN[SRCDIST+nprow]; |
| 278 |
/*
|
| 279 |
* Spread U
|
| 280 |
*/
|
| 281 |
do
|
| 282 |
{
|
| 283 |
mask ^= ip2; |
| 284 |
|
| 285 |
if( ( mydist & mask ) == 0 ) |
| 286 |
{
|
| 287 |
k = il + ip2; ibuf = IPLEN[SRCDIST+il]; |
| 288 |
lbuf = ( k >= nprow ? lgth : IPLEN[SRCDIST+k] ) - ibuf; |
| 289 |
|
| 290 |
if( lbuf > 0 ) |
| 291 |
{
|
| 292 |
partner = mydist ^ ip2; |
| 293 |
|
| 294 |
if( mydist & ip2 )
|
| 295 |
{
|
| 296 |
#if 0
|
| 297 |
if( ierr == MPI_SUCCESS )
|
| 298 |
{
|
| 299 |
if( LDU == N )
|
| 300 |
ierr = MPI_Type_contiguous( lbuf*LDU, MPI_DOUBLE,
|
| 301 |
&type );
|
| 302 |
else
|
| 303 |
ierr = MPI_Type_vector( lbuf, N, LDU, MPI_DOUBLE,
|
| 304 |
&type );
|
| 305 |
}
|
| 306 |
if( ierr == MPI_SUCCESS )
|
| 307 |
ierr = MPI_Type_commit( &type );
|
| 308 |
if( ierr == MPI_SUCCESS )
|
| 309 |
ierr = MPI_Recv( Mptr( U, 0, ibuf, LDU ), 1, type,
|
| 310 |
IPMAP[SRCDIST+partner], Cmsgid,
|
| 311 |
comm, &status );
|
| 312 |
if( ierr == MPI_SUCCESS )
|
| 313 |
ierr = MPI_Type_free( &type );
|
| 314 |
#else
|
| 315 |
/*
|
| 316 |
* In our case, LDU is N - do not use the MPI Datatypes
|
| 317 |
*/
|
| 318 |
if( ierr == MPI_SUCCESS )
|
| 319 |
ierr = MPI_Recv( Mptr( U, 0, ibuf, LDU ), lbuf*N,
|
| 320 |
MPI_DOUBLE, IPMAP[SRCDIST+partner], |
| 321 |
Cmsgid, comm, &status ); |
| 322 |
#endif
|
| 323 |
} |
| 324 |
else if( partner < nprow ) |
| 325 |
{
|
| 326 |
#if 0
|
| 327 |
if( ierr == MPI_SUCCESS )
|
| 328 |
{
|
| 329 |
if( LDU == N )
|
| 330 |
ierr = MPI_Type_contiguous( lbuf*LDU, MPI_DOUBLE,
|
| 331 |
&type );
|
| 332 |
else
|
| 333 |
ierr = MPI_Type_vector( lbuf, N, LDU, MPI_DOUBLE,
|
| 334 |
&type );
|
| 335 |
}
|
| 336 |
if( ierr == MPI_SUCCESS )
|
| 337 |
ierr = MPI_Type_commit( &type );
|
| 338 |
if( ierr == MPI_SUCCESS )
|
| 339 |
ierr = MPI_Send( Mptr( U, 0, ibuf, LDU ), 1, type,
|
| 340 |
IPMAP[SRCDIST+partner], Cmsgid,
|
| 341 |
comm );
|
| 342 |
if( ierr == MPI_SUCCESS )
|
| 343 |
ierr = MPI_Type_free( &type );
|
| 344 |
#else
|
| 345 |
/*
|
| 346 |
* In our case, LDU is N - do not use the MPI Datatypes
|
| 347 |
*/
|
| 348 |
if( ierr == MPI_SUCCESS )
|
| 349 |
ierr = MPI_Send( Mptr( U, 0, ibuf, LDU ), lbuf*N,
|
| 350 |
MPI_DOUBLE, IPMAP[SRCDIST+partner], |
| 351 |
Cmsgid, comm ); |
| 352 |
#endif
|
| 353 |
} |
| 354 |
} |
| 355 |
} |
| 356 |
|
| 357 |
if( mydist2 < ip2 ) { ip2 >>= 1; il -= ip2; } |
| 358 |
else { mydist2 -= ip2; ip2 >>= 1; il += ip2; } |
| 359 |
/*
|
| 360 |
* Probe for column panel - forward it when available
|
| 361 |
*/
|
| 362 |
if( *IFLAG == HPL_KEEP_TESTING ) (void) HPL_bcast( PBCST, IFLAG ); |
| 363 |
|
| 364 |
} while( ip2 > 0 ); |
| 365 |
} |
| 366 |
|
| 367 |
if( ierr != MPI_SUCCESS )
|
| 368 |
{ HPL_pabort( __LINE__, "HPL_spreadT", "MPI call failed" ); }
|
| 369 |
/*
|
| 370 |
* End of HPL_spreadT
|
| 371 |
*/
|
| 372 |
} |