root / testing / ptest / HPL_pdinfo.c
Historique | Voir | Annoter | Télécharger (43,51 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_pdinfo
|
| 54 |
( |
| 55 |
HPL_T_test * TEST, |
| 56 |
int * NS,
|
| 57 |
int * N,
|
| 58 |
int * NBS,
|
| 59 |
int * NB,
|
| 60 |
HPL_T_ORDER * PMAPPIN, |
| 61 |
int * NPQS,
|
| 62 |
int * P,
|
| 63 |
int * Q,
|
| 64 |
int * NPFS,
|
| 65 |
HPL_T_FACT * PF, |
| 66 |
int * NBMS,
|
| 67 |
int * NBM,
|
| 68 |
int * NDVS,
|
| 69 |
int * NDV,
|
| 70 |
int * NRFS,
|
| 71 |
HPL_T_FACT * RF, |
| 72 |
int * NTPS,
|
| 73 |
HPL_T_TOP * TP, |
| 74 |
int * NDHS,
|
| 75 |
int * DH,
|
| 76 |
HPL_T_SWAP * FSWAP, |
| 77 |
int * TSWAP,
|
| 78 |
int * L1NOTRAN,
|
| 79 |
int * UNOTRAN,
|
| 80 |
int * EQUIL,
|
| 81 |
int * ALIGN
|
| 82 |
) |
| 83 |
#else
|
| 84 |
void HPL_pdinfo
|
| 85 |
( TEST, NS, N, NBS, NB, PMAPPIN, NPQS, P, Q, NPFS, PF, NBMS, NBM, NDVS, NDV, NRFS, RF, NTPS, TP, NDHS, DH, FSWAP, TSWAP, L1NOTRAN, UNOTRAN, EQUIL, ALIGN ) |
| 86 |
HPL_T_test * TEST; |
| 87 |
int * NS;
|
| 88 |
int * N;
|
| 89 |
int * NBS;
|
| 90 |
int * NB;
|
| 91 |
HPL_T_ORDER * PMAPPIN; |
| 92 |
int * NPQS;
|
| 93 |
int * P;
|
| 94 |
int * Q;
|
| 95 |
int * NPFS;
|
| 96 |
HPL_T_FACT * PF; |
| 97 |
int * NBMS;
|
| 98 |
int * NBM;
|
| 99 |
int * NDVS;
|
| 100 |
int * NDV;
|
| 101 |
int * NRFS;
|
| 102 |
HPL_T_FACT * RF; |
| 103 |
int * NTPS;
|
| 104 |
HPL_T_TOP * TP; |
| 105 |
int * NDHS;
|
| 106 |
int * DH;
|
| 107 |
HPL_T_SWAP * FSWAP; |
| 108 |
int * TSWAP;
|
| 109 |
int * L1NOTRAN;
|
| 110 |
int * UNOTRAN;
|
| 111 |
int * EQUIL;
|
| 112 |
int * ALIGN;
|
| 113 |
#endif
|
| 114 |
{
|
| 115 |
/*
|
| 116 |
* Purpose
|
| 117 |
* =======
|
| 118 |
*
|
| 119 |
* HPL_pdinfo reads the startup information for the various tests and
|
| 120 |
* transmits it to all processes.
|
| 121 |
*
|
| 122 |
* Arguments
|
| 123 |
* =========
|
| 124 |
*
|
| 125 |
* TEST (global output) HPL_T_test *
|
| 126 |
* On entry, TEST points to a testing data structure. On exit,
|
| 127 |
* the fields of this data structure are initialized as follows:
|
| 128 |
* TEST->outfp specifies the output file where the results will
|
| 129 |
* be printed. It is only defined and used by the process 0 of
|
| 130 |
* the grid. TEST->thrsh specifies the threshhold value for the
|
| 131 |
* test ratio. TEST->epsil is the relative machine precision of
|
| 132 |
* the distributed computer. Finally the test counters, kfail,
|
| 133 |
* kpass, kskip, ktest are initialized to zero.
|
| 134 |
*
|
| 135 |
* NS (global output) int *
|
| 136 |
* On exit, NS specifies the number of different problem sizes
|
| 137 |
* to be tested. NS is less than or equal to HPL_MAX_PARAM.
|
| 138 |
*
|
| 139 |
* N (global output) int *
|
| 140 |
* On entry, N is an array of dimension HPL_MAX_PARAM. On exit,
|
| 141 |
* the first NS entries of this array contain the problem sizes
|
| 142 |
* to run the code with.
|
| 143 |
*
|
| 144 |
* NBS (global output) int *
|
| 145 |
* On exit, NBS specifies the number of different distribution
|
| 146 |
* blocking factors to be tested. NBS must be less than or equal
|
| 147 |
* to HPL_MAX_PARAM.
|
| 148 |
*
|
| 149 |
* NB (global output) int *
|
| 150 |
* On exit, PMAPPIN specifies the process mapping onto the no-
|
| 151 |
* des of the MPI machine configuration. PMAPPIN defaults to
|
| 152 |
* row-major ordering.
|
| 153 |
*
|
| 154 |
* PMAPPIN (global output) HPL_T_ORDER *
|
| 155 |
* On entry, NB is an array of dimension HPL_MAX_PARAM. On exit,
|
| 156 |
* the first NBS entries of this array contain the values of the
|
| 157 |
* various distribution blocking factors, to run the code with.
|
| 158 |
*
|
| 159 |
* NPQS (global output) int *
|
| 160 |
* On exit, NPQS specifies the number of different values that
|
| 161 |
* can be used for P and Q, i.e., the number of process grids to
|
| 162 |
* run the code with. NPQS must be less than or equal to
|
| 163 |
* HPL_MAX_PARAM.
|
| 164 |
*
|
| 165 |
* P (global output) int *
|
| 166 |
* On entry, P is an array of dimension HPL_MAX_PARAM. On exit,
|
| 167 |
* the first NPQS entries of this array contain the values of P,
|
| 168 |
* the number of process rows of the NPQS grids to run the code
|
| 169 |
* with.
|
| 170 |
*
|
| 171 |
* Q (global output) int *
|
| 172 |
* On entry, Q is an array of dimension HPL_MAX_PARAM. On exit,
|
| 173 |
* the first NPQS entries of this array contain the values of Q,
|
| 174 |
* the number of process columns of the NPQS grids to run the
|
| 175 |
* code with.
|
| 176 |
*
|
| 177 |
* NPFS (global output) int *
|
| 178 |
* On exit, NPFS specifies the number of different values that
|
| 179 |
* can be used for PF : the panel factorization algorithm to run
|
| 180 |
* the code with. NPFS is less than or equal to HPL_MAX_PARAM.
|
| 181 |
*
|
| 182 |
* PF (global output) HPL_T_FACT *
|
| 183 |
* On entry, PF is an array of dimension HPL_MAX_PARAM. On exit,
|
| 184 |
* the first NPFS entries of this array contain the various
|
| 185 |
* panel factorization algorithms to run the code with.
|
| 186 |
*
|
| 187 |
* NBMS (global output) int *
|
| 188 |
* On exit, NBMS specifies the number of various recursive
|
| 189 |
* stopping criteria to be tested. NBMS must be less than or
|
| 190 |
* equal to HPL_MAX_PARAM.
|
| 191 |
*
|
| 192 |
* NBM (global output) int *
|
| 193 |
* On entry, NBM is an array of dimension HPL_MAX_PARAM. On
|
| 194 |
* exit, the first NBMS entries of this array contain the values
|
| 195 |
* of the various recursive stopping criteria to be tested.
|
| 196 |
*
|
| 197 |
* NDVS (global output) int *
|
| 198 |
* On exit, NDVS specifies the number of various numbers of
|
| 199 |
* panels in recursion to be tested. NDVS is less than or equal
|
| 200 |
* to HPL_MAX_PARAM.
|
| 201 |
*
|
| 202 |
* NDV (global output) int *
|
| 203 |
* On entry, NDV is an array of dimension HPL_MAX_PARAM. On
|
| 204 |
* exit, the first NDVS entries of this array contain the values
|
| 205 |
* of the various numbers of panels in recursion to be tested.
|
| 206 |
*
|
| 207 |
* NRFS (global output) int *
|
| 208 |
* On exit, NRFS specifies the number of different values that
|
| 209 |
* can be used for RF : the recursive factorization algorithm to
|
| 210 |
* be tested. NRFS is less than or equal to HPL_MAX_PARAM.
|
| 211 |
*
|
| 212 |
* RF (global output) HPL_T_FACT *
|
| 213 |
* On entry, RF is an array of dimension HPL_MAX_PARAM. On exit,
|
| 214 |
* the first NRFS entries of this array contain the various
|
| 215 |
* recursive factorization algorithms to run the code with.
|
| 216 |
*
|
| 217 |
* NTPS (global output) int *
|
| 218 |
* On exit, NTPS specifies the number of different values that
|
| 219 |
* can be used for the broadcast topologies to be tested. NTPS
|
| 220 |
* is less than or equal to HPL_MAX_PARAM.
|
| 221 |
*
|
| 222 |
* TP (global output) HPL_T_TOP *
|
| 223 |
* On entry, TP is an array of dimension HPL_MAX_PARAM. On exit,
|
| 224 |
* the first NTPS entries of this array contain the various
|
| 225 |
* broadcast (along rows) topologies to run the code with.
|
| 226 |
*
|
| 227 |
* NDHS (global output) int *
|
| 228 |
* On exit, NDHS specifies the number of different values that
|
| 229 |
* can be used for the lookahead depths to be tested. NDHS is
|
| 230 |
* less than or equal to HPL_MAX_PARAM.
|
| 231 |
*
|
| 232 |
* DH (global output) int *
|
| 233 |
* On entry, DH is an array of dimension HPL_MAX_PARAM. On
|
| 234 |
* exit, the first NDHS entries of this array contain the values
|
| 235 |
* of lookahead depths to run the code with. Such a value is at
|
| 236 |
* least 0 (no-lookahead) or greater than zero.
|
| 237 |
*
|
| 238 |
* FSWAP (global output) HPL_T_SWAP *
|
| 239 |
* On exit, FSWAP specifies the swapping algorithm to be used in
|
| 240 |
* all tests.
|
| 241 |
*
|
| 242 |
* TSWAP (global output) int *
|
| 243 |
* On exit, TSWAP specifies the swapping threshold as a number
|
| 244 |
* of columns when the mixed swapping algorithm was chosen.
|
| 245 |
*
|
| 246 |
* L1NOTRA (global output) int *
|
| 247 |
* On exit, L1NOTRAN specifies whether the upper triangle of the
|
| 248 |
* panels of columns should be stored in no-transposed form
|
| 249 |
* (L1NOTRAN=1) or in transposed form (L1NOTRAN=0).
|
| 250 |
*
|
| 251 |
* UNOTRAN (global output) int *
|
| 252 |
* On exit, UNOTRAN specifies whether the panels of rows should
|
| 253 |
* be stored in no-transposed form (UNOTRAN=1) or transposed
|
| 254 |
* form (UNOTRAN=0) during their broadcast.
|
| 255 |
*
|
| 256 |
* EQUIL (global output) int *
|
| 257 |
* On exit, EQUIL specifies whether equilibration during the
|
| 258 |
* swap-broadcast of the panel of rows should be performed
|
| 259 |
* (EQUIL=1) or not (EQUIL=0).
|
| 260 |
*
|
| 261 |
* ALIGN (global output) int *
|
| 262 |
* On exit, ALIGN specifies the alignment of the dynamically
|
| 263 |
* allocated buffers in double precision words. ALIGN is greater
|
| 264 |
* than zero.
|
| 265 |
*
|
| 266 |
* ---------------------------------------------------------------------
|
| 267 |
*/
|
| 268 |
/*
|
| 269 |
* .. Local Variables ..
|
| 270 |
*/
|
| 271 |
char file[HPL_LINE_MAX], line[HPL_LINE_MAX],
|
| 272 |
auth[HPL_LINE_MAX], num [HPL_LINE_MAX]; |
| 273 |
FILE * infp; |
| 274 |
int * iwork = NULL; |
| 275 |
char * lineptr;
|
| 276 |
int error=0, fid, i, j, lwork, maxp, nprocs, |
| 277 |
rank, size; |
| 278 |
/* ..
|
| 279 |
* .. Executable Statements ..
|
| 280 |
*/
|
| 281 |
MPI_Comm_rank( MPI_COMM_WORLD, &rank ); |
| 282 |
MPI_Comm_size( MPI_COMM_WORLD, &size ); |
| 283 |
/*
|
| 284 |
* Initialize the TEST data structure with default values
|
| 285 |
*/
|
| 286 |
TEST->outfp = stderr; TEST->epsil = 2.0e-16; TEST->thrsh = 16.0; |
| 287 |
TEST->kfail = TEST->kpass = TEST->kskip = TEST->ktest = 0;
|
| 288 |
/*
|
| 289 |
* Process 0 reads the input data, broadcasts to other processes and
|
| 290 |
* writes needed information to TEST->outfp.
|
| 291 |
*/
|
| 292 |
if( rank == 0 ) |
| 293 |
{
|
| 294 |
/*
|
| 295 |
* Open file and skip data file header
|
| 296 |
*/
|
| 297 |
if( ( infp = fopen( "HPL.dat", "r" ) ) == NULL ) |
| 298 |
{
|
| 299 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo",
|
| 300 |
"cannot open file HPL.dat" );
|
| 301 |
error = 1; goto label_error; |
| 302 |
} |
| 303 |
|
| 304 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 305 |
(void) fgets( auth, HPL_LINE_MAX - 2, infp ); |
| 306 |
/*
|
| 307 |
* Read name and unit number for summary output file
|
| 308 |
*/
|
| 309 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 310 |
(void) sscanf( line, "%s", file ); |
| 311 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 312 |
(void) sscanf( line, "%s", num ); |
| 313 |
fid = atoi( num ); |
| 314 |
if ( fid == 6 ) TEST->outfp = stdout; |
| 315 |
else if( fid == 7 ) TEST->outfp = stderr; |
| 316 |
else if( ( TEST->outfp = fopen( file, "w" ) ) == NULL ) |
| 317 |
{
|
| 318 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo", "cannot open file %s.", |
| 319 |
file ); |
| 320 |
error = 1; goto label_error; |
| 321 |
} |
| 322 |
/*
|
| 323 |
* Read and check the parameter values for the tests.
|
| 324 |
*
|
| 325 |
* Problem size (>=0) (N)
|
| 326 |
*/
|
| 327 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 328 |
(void) sscanf( line, "%s", num ); *NS = atoi( num ); |
| 329 |
if( ( *NS < 1 ) || ( *NS > HPL_MAX_PARAM ) ) |
| 330 |
{
|
| 331 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo", "%s %d", |
| 332 |
"Number of values of N is less than 1 or greater than",
|
| 333 |
HPL_MAX_PARAM ); |
| 334 |
error = 1; goto label_error; |
| 335 |
} |
| 336 |
|
| 337 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); lineptr = line; |
| 338 |
for( i = 0; i < *NS; i++ ) |
| 339 |
{
|
| 340 |
(void) sscanf( lineptr, "%s", num ); lineptr += strlen( num ) + 1; |
| 341 |
if( ( N[ i ] = atoi( num ) ) < 0 ) |
| 342 |
{
|
| 343 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo",
|
| 344 |
"Value of N less than 0" );
|
| 345 |
error = 1; goto label_error; |
| 346 |
} |
| 347 |
} |
| 348 |
/*
|
| 349 |
* Block size (>=1) (NB)
|
| 350 |
*/
|
| 351 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 352 |
(void) sscanf( line, "%s", num ); *NBS = atoi( num ); |
| 353 |
if( ( *NBS < 1 ) || ( *NBS > HPL_MAX_PARAM ) ) |
| 354 |
{
|
| 355 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo", "%s %s %d", |
| 356 |
"Number of values of NB is less than 1 or",
|
| 357 |
"greater than", HPL_MAX_PARAM );
|
| 358 |
error = 1; goto label_error; |
| 359 |
} |
| 360 |
|
| 361 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); lineptr = line; |
| 362 |
for( i = 0; i < *NBS; i++ ) |
| 363 |
{
|
| 364 |
(void) sscanf( lineptr, "%s", num ); lineptr += strlen( num ) + 1; |
| 365 |
if( ( NB[ i ] = atoi( num ) ) < 1 ) |
| 366 |
{
|
| 367 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo",
|
| 368 |
"Value of NB less than 1" );
|
| 369 |
error = 1; goto label_error; |
| 370 |
} |
| 371 |
} |
| 372 |
/*
|
| 373 |
* Process grids, mapping, (>=1) (P, Q)
|
| 374 |
*/
|
| 375 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 376 |
(void) sscanf( line, "%s", num ); |
| 377 |
*PMAPPIN = ( atoi( num ) == 1 ? HPL_COLUMN_MAJOR : HPL_ROW_MAJOR );
|
| 378 |
|
| 379 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 380 |
(void) sscanf( line, "%s", num ); *NPQS = atoi( num ); |
| 381 |
if( ( *NPQS < 1 ) || ( *NPQS > HPL_MAX_PARAM ) ) |
| 382 |
{
|
| 383 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo", "%s %s %d", |
| 384 |
"Number of values of grids is less",
|
| 385 |
"than 1 or greater than", HPL_MAX_PARAM );
|
| 386 |
error = 1; goto label_error; |
| 387 |
} |
| 388 |
|
| 389 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); lineptr = line; |
| 390 |
for( i = 0; i < *NPQS; i++ ) |
| 391 |
{
|
| 392 |
(void) sscanf( lineptr, "%s", num ); lineptr += strlen( num ) + 1; |
| 393 |
if( ( P[ i ] = atoi( num ) ) < 1 ) |
| 394 |
{
|
| 395 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo",
|
| 396 |
"Value of P less than 1" );
|
| 397 |
error = 1; goto label_error; |
| 398 |
} |
| 399 |
} |
| 400 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); lineptr = line; |
| 401 |
for( i = 0; i < *NPQS; i++ ) |
| 402 |
{
|
| 403 |
(void) sscanf( lineptr, "%s", num ); lineptr += strlen( num ) + 1; |
| 404 |
if( ( Q[ i ] = atoi( num ) ) < 1 ) |
| 405 |
{
|
| 406 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo",
|
| 407 |
"Value of Q less than 1" );
|
| 408 |
error = 1; goto label_error; |
| 409 |
} |
| 410 |
} |
| 411 |
/*
|
| 412 |
* Check for enough processes in machine configuration
|
| 413 |
*/
|
| 414 |
maxp = 0;
|
| 415 |
for( i = 0; i < *NPQS; i++ ) |
| 416 |
{ nprocs = P[i] * Q[i]; maxp = Mmax( maxp, nprocs ); }
|
| 417 |
if( maxp > size )
|
| 418 |
{
|
| 419 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo",
|
| 420 |
"Need at least %d processes for these tests", maxp );
|
| 421 |
error = 1; goto label_error; |
| 422 |
} |
| 423 |
/*
|
| 424 |
* Checking threshold value (TEST->thrsh)
|
| 425 |
*/
|
| 426 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 427 |
(void) sscanf( line, "%s", num ); TEST->thrsh = atof( num ); |
| 428 |
/*
|
| 429 |
* Panel factorization algorithm (PF)
|
| 430 |
*/
|
| 431 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 432 |
(void) sscanf( line, "%s", num ); *NPFS = atoi( num ); |
| 433 |
if( ( *NPFS < 1 ) || ( *NPFS > HPL_MAX_PARAM ) ) |
| 434 |
{
|
| 435 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo", "%s %s %d", |
| 436 |
"number of values of PFACT",
|
| 437 |
"is less than 1 or greater than", HPL_MAX_PARAM );
|
| 438 |
error = 1; goto label_error; |
| 439 |
} |
| 440 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); lineptr = line; |
| 441 |
for( i = 0; i < *NPFS; i++ ) |
| 442 |
{
|
| 443 |
(void) sscanf( lineptr, "%s", num ); lineptr += strlen( num ) + 1; |
| 444 |
j = atoi( num ); |
| 445 |
if( j == 0 ) PF[ i ] = HPL_LEFT_LOOKING; |
| 446 |
else if( j == 1 ) PF[ i ] = HPL_CROUT; |
| 447 |
else if( j == 2 ) PF[ i ] = HPL_RIGHT_LOOKING; |
| 448 |
else PF[ i ] = HPL_RIGHT_LOOKING;
|
| 449 |
} |
| 450 |
/*
|
| 451 |
* Recursive stopping criterium (>=1) (NBM)
|
| 452 |
*/
|
| 453 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 454 |
(void) sscanf( line, "%s", num ); *NBMS = atoi( num ); |
| 455 |
if( ( *NBMS < 1 ) || ( *NBMS > HPL_MAX_PARAM ) ) |
| 456 |
{
|
| 457 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo", "%s %s %d", |
| 458 |
"Number of values of NBMIN",
|
| 459 |
"is less than 1 or greater than", HPL_MAX_PARAM );
|
| 460 |
error = 1; goto label_error; |
| 461 |
} |
| 462 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); lineptr = line; |
| 463 |
for( i = 0; i < *NBMS; i++ ) |
| 464 |
{
|
| 465 |
(void) sscanf( lineptr, "%s", num ); lineptr += strlen( num ) + 1; |
| 466 |
if( ( NBM[ i ] = atoi( num ) ) < 1 ) |
| 467 |
{
|
| 468 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo",
|
| 469 |
"Value of NBMIN less than 1" );
|
| 470 |
error = 1; goto label_error; |
| 471 |
} |
| 472 |
} |
| 473 |
/*
|
| 474 |
* Number of panels in recursion (>=2) (NDV)
|
| 475 |
*/
|
| 476 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 477 |
(void) sscanf( line, "%s", num ); *NDVS = atoi( num ); |
| 478 |
if( ( *NDVS < 1 ) || ( *NDVS > HPL_MAX_PARAM ) ) |
| 479 |
{
|
| 480 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo", "%s %s %d", |
| 481 |
"Number of values of NDIV",
|
| 482 |
"is less than 1 or greater than", HPL_MAX_PARAM );
|
| 483 |
error = 1; goto label_error; |
| 484 |
} |
| 485 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); lineptr = line; |
| 486 |
for( i = 0; i < *NDVS; i++ ) |
| 487 |
{
|
| 488 |
(void) sscanf( lineptr, "%s", num ); lineptr += strlen( num ) + 1; |
| 489 |
if( ( NDV[ i ] = atoi( num ) ) < 2 ) |
| 490 |
{
|
| 491 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo",
|
| 492 |
"Value of NDIV less than 2" );
|
| 493 |
error = 1; goto label_error; |
| 494 |
} |
| 495 |
} |
| 496 |
/*
|
| 497 |
* Recursive panel factorization (RF)
|
| 498 |
*/
|
| 499 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 500 |
(void) sscanf( line, "%s", num ); *NRFS = atoi( num ); |
| 501 |
if( ( *NRFS < 1 ) || ( *NRFS > HPL_MAX_PARAM ) ) |
| 502 |
{
|
| 503 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo", "%s %s %d", |
| 504 |
"Number of values of RFACT",
|
| 505 |
"is less than 1 or greater than", HPL_MAX_PARAM );
|
| 506 |
error = 1; goto label_error; |
| 507 |
} |
| 508 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); lineptr = line; |
| 509 |
for( i = 0; i < *NRFS; i++ ) |
| 510 |
{
|
| 511 |
(void) sscanf( lineptr, "%s", num ); lineptr += strlen( num ) + 1; |
| 512 |
j = atoi( num ); |
| 513 |
if( j == 0 ) RF[ i ] = HPL_LEFT_LOOKING; |
| 514 |
else if( j == 1 ) RF[ i ] = HPL_CROUT; |
| 515 |
else if( j == 2 ) RF[ i ] = HPL_RIGHT_LOOKING; |
| 516 |
else RF[ i ] = HPL_RIGHT_LOOKING;
|
| 517 |
} |
| 518 |
/*
|
| 519 |
* Broadcast topology (TP) (0=rg, 1=2rg, 2=rgM, 3=2rgM, 4=L)
|
| 520 |
*/
|
| 521 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 522 |
(void) sscanf( line, "%s", num ); *NTPS = atoi( num ); |
| 523 |
if( ( *NTPS < 1 ) || ( *NTPS > HPL_MAX_PARAM ) ) |
| 524 |
{
|
| 525 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo", "%s %s %d", |
| 526 |
"Number of values of BCAST",
|
| 527 |
"is less than 1 or greater than", HPL_MAX_PARAM );
|
| 528 |
error = 1; goto label_error; |
| 529 |
} |
| 530 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); lineptr = line; |
| 531 |
for( i = 0; i < *NTPS; i++ ) |
| 532 |
{
|
| 533 |
(void) sscanf( lineptr, "%s", num ); lineptr += strlen( num ) + 1; |
| 534 |
j = atoi( num ); |
| 535 |
if( j == 0 ) TP[ i ] = HPL_1RING; |
| 536 |
else if( j == 1 ) TP[ i ] = HPL_1RING_M; |
| 537 |
else if( j == 2 ) TP[ i ] = HPL_2RING; |
| 538 |
else if( j == 3 ) TP[ i ] = HPL_2RING_M; |
| 539 |
else if( j == 4 ) TP[ i ] = HPL_BLONG; |
| 540 |
else if( j == 5 ) TP[ i ] = HPL_BLONG_M; |
| 541 |
else TP[ i ] = HPL_1RING_M;
|
| 542 |
} |
| 543 |
/*
|
| 544 |
* Lookahead depth (>=0) (NDH)
|
| 545 |
*/
|
| 546 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 547 |
(void) sscanf( line, "%s", num ); *NDHS = atoi( num ); |
| 548 |
if( ( *NDHS < 1 ) || ( *NDHS > HPL_MAX_PARAM ) ) |
| 549 |
{
|
| 550 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo", "%s %s %d", |
| 551 |
"Number of values of DEPTH",
|
| 552 |
"is less than 1 or greater than", HPL_MAX_PARAM );
|
| 553 |
error = 1; goto label_error; |
| 554 |
} |
| 555 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); lineptr = line; |
| 556 |
for( i = 0; i < *NDHS; i++ ) |
| 557 |
{
|
| 558 |
(void) sscanf( lineptr, "%s", num ); |
| 559 |
lineptr += strlen( num ) + 1;
|
| 560 |
if( ( DH[ i ] = atoi( num ) ) < 0 ) |
| 561 |
{
|
| 562 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo",
|
| 563 |
"Value of DEPTH less than 0" );
|
| 564 |
error = 1; goto label_error; |
| 565 |
} |
| 566 |
} |
| 567 |
/*
|
| 568 |
* Swapping algorithm (0,1 or 2) (FSWAP)
|
| 569 |
*/
|
| 570 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 571 |
(void) sscanf( line, "%s", num ); j = atoi( num ); |
| 572 |
if( j == 0 ) *FSWAP = HPL_SWAP00; |
| 573 |
else if( j == 1 ) *FSWAP = HPL_SWAP01; |
| 574 |
else if( j == 2 ) *FSWAP = HPL_SW_MIX; |
| 575 |
else *FSWAP = HPL_SWAP01;
|
| 576 |
/*
|
| 577 |
* Swapping threshold (>=0) (TSWAP)
|
| 578 |
*/
|
| 579 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 580 |
(void) sscanf( line, "%s", num ); *TSWAP = atoi( num ); |
| 581 |
if( *TSWAP <= 0 ) *TSWAP = 0; |
| 582 |
/*
|
| 583 |
* L1 in (no-)transposed form (0 or 1)
|
| 584 |
*/
|
| 585 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 586 |
(void) sscanf( line, "%s", num ); *L1NOTRAN = atoi( num ); |
| 587 |
if( ( *L1NOTRAN != 0 ) && ( *L1NOTRAN != 1 ) ) *L1NOTRAN = 0; |
| 588 |
/*
|
| 589 |
* U in (no-)transposed form (0 or 1)
|
| 590 |
*/
|
| 591 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 592 |
(void) sscanf( line, "%s", num ); *UNOTRAN = atoi( num ); |
| 593 |
if( ( *UNOTRAN != 0 ) && ( *UNOTRAN != 1 ) ) *UNOTRAN = 0; |
| 594 |
/*
|
| 595 |
* Equilibration (0=no, 1=yes)
|
| 596 |
*/
|
| 597 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 598 |
(void) sscanf( line, "%s", num ); *EQUIL = atoi( num ); |
| 599 |
if( ( *EQUIL != 0 ) && ( *EQUIL != 1 ) ) *EQUIL = 1; |
| 600 |
/*
|
| 601 |
* Memory alignment in bytes (> 0) (ALIGN)
|
| 602 |
*/
|
| 603 |
(void) fgets( line, HPL_LINE_MAX - 2, infp ); |
| 604 |
(void) sscanf( line, "%s", num ); *ALIGN = atoi( num ); |
| 605 |
if( *ALIGN <= 0 ) *ALIGN = 4; |
| 606 |
/*
|
| 607 |
* Close input file
|
| 608 |
*/
|
| 609 |
label_error:
|
| 610 |
(void) fclose( infp );
|
| 611 |
} |
| 612 |
else { TEST->outfp = NULL; } |
| 613 |
/*
|
| 614 |
* Check for error on reading input file
|
| 615 |
*/
|
| 616 |
(void) HPL_all_reduce( (void *)(&error), 1, HPL_INT, HPL_max, |
| 617 |
MPI_COMM_WORLD ); |
| 618 |
if( error )
|
| 619 |
{
|
| 620 |
if( rank == 0 ) |
| 621 |
HPL_pwarn( stderr, __LINE__, "HPL_pdinfo",
|
| 622 |
"Illegal input in file HPL.dat. Exiting ..." );
|
| 623 |
MPI_Finalize(); |
| 624 |
#ifdef HPL_CALL_VSIPL
|
| 625 |
(void) vsip_finalize( NULL ); |
| 626 |
#endif
|
| 627 |
exit( 1 );
|
| 628 |
} |
| 629 |
/*
|
| 630 |
* Compute and broadcast machine epsilon
|
| 631 |
*/
|
| 632 |
TEST->epsil = HPL_pdlamch( MPI_COMM_WORLD, HPL_MACH_EPS ); |
| 633 |
/*
|
| 634 |
* Pack information arrays and broadcast
|
| 635 |
*/
|
| 636 |
(void) HPL_broadcast( (void *)(&(TEST->thrsh)), 1, HPL_DOUBLE, 0, |
| 637 |
MPI_COMM_WORLD ); |
| 638 |
/*
|
| 639 |
* Broadcast array sizes
|
| 640 |
*/
|
| 641 |
iwork = (int *)malloc( (size_t)(15) * sizeof( int ) ); |
| 642 |
if( rank == 0 ) |
| 643 |
{
|
| 644 |
iwork[ 0] = *NS; iwork[ 1] = *NBS; |
| 645 |
iwork[ 2] = ( *PMAPPIN == HPL_ROW_MAJOR ? 0 : 1 ); |
| 646 |
iwork[ 3] = *NPQS; iwork[ 4] = *NPFS; iwork[ 5] = *NBMS; |
| 647 |
iwork[ 6] = *NDVS; iwork[ 7] = *NRFS; iwork[ 8] = *NTPS; |
| 648 |
iwork[ 9] = *NDHS; iwork[10] = *TSWAP; iwork[11] = *L1NOTRAN; |
| 649 |
iwork[12] = *UNOTRAN; iwork[13] = *EQUIL; iwork[14] = *ALIGN; |
| 650 |
} |
| 651 |
(void) HPL_broadcast( (void *)iwork, 15, HPL_INT, 0, MPI_COMM_WORLD ); |
| 652 |
if( rank != 0 ) |
| 653 |
{
|
| 654 |
*NS = iwork[ 0]; *NBS = iwork[ 1]; |
| 655 |
*PMAPPIN = ( iwork[ 2] == 0 ? HPL_ROW_MAJOR : HPL_COLUMN_MAJOR ); |
| 656 |
*NPQS = iwork[ 3]; *NPFS = iwork[ 4]; *NBMS = iwork[ 5]; |
| 657 |
*NDVS = iwork[ 6]; *NRFS = iwork[ 7]; *NTPS = iwork[ 8]; |
| 658 |
*NDHS = iwork[ 9]; *TSWAP = iwork[10]; *L1NOTRAN = iwork[11]; |
| 659 |
*UNOTRAN = iwork[12]; *EQUIL = iwork[13]; *ALIGN = iwork[14]; |
| 660 |
} |
| 661 |
if( iwork ) free( iwork );
|
| 662 |
/*
|
| 663 |
* Pack information arrays and broadcast
|
| 664 |
*/
|
| 665 |
lwork = (*NS) + (*NBS) + 2 * (*NPQS) + (*NPFS) + (*NBMS) +
|
| 666 |
(*NDVS) + (*NRFS) + (*NTPS) + (*NDHS) + 1;
|
| 667 |
iwork = (int *)malloc( (size_t)(lwork) * sizeof( int ) ); |
| 668 |
if( rank == 0 ) |
| 669 |
{
|
| 670 |
j = 0;
|
| 671 |
for( i = 0; i < *NS; i++ ) { iwork[j] = N [i]; j++; } |
| 672 |
for( i = 0; i < *NBS; i++ ) { iwork[j] = NB[i]; j++; } |
| 673 |
for( i = 0; i < *NPQS; i++ ) { iwork[j] = P [i]; j++; } |
| 674 |
for( i = 0; i < *NPQS; i++ ) { iwork[j] = Q [i]; j++; } |
| 675 |
for( i = 0; i < *NPFS; i++ ) |
| 676 |
{
|
| 677 |
if( PF[i] == HPL_LEFT_LOOKING ) iwork[j] = 0; |
| 678 |
else if( PF[i] == HPL_CROUT ) iwork[j] = 1; |
| 679 |
else if( PF[i] == HPL_RIGHT_LOOKING ) iwork[j] = 2; |
| 680 |
j++; |
| 681 |
} |
| 682 |
for( i = 0; i < *NBMS; i++ ) { iwork[j] = NBM[i]; j++; } |
| 683 |
for( i = 0; i < *NDVS; i++ ) { iwork[j] = NDV[i]; j++; } |
| 684 |
for( i = 0; i < *NRFS; i++ ) |
| 685 |
{
|
| 686 |
if( RF[i] == HPL_LEFT_LOOKING ) iwork[j] = 0; |
| 687 |
else if( RF[i] == HPL_CROUT ) iwork[j] = 1; |
| 688 |
else if( RF[i] == HPL_RIGHT_LOOKING ) iwork[j] = 2; |
| 689 |
j++; |
| 690 |
} |
| 691 |
for( i = 0; i < *NTPS; i++ ) |
| 692 |
{
|
| 693 |
if( TP[i] == HPL_1RING ) iwork[j] = 0; |
| 694 |
else if( TP[i] == HPL_1RING_M ) iwork[j] = 1; |
| 695 |
else if( TP[i] == HPL_2RING ) iwork[j] = 2; |
| 696 |
else if( TP[i] == HPL_2RING_M ) iwork[j] = 3; |
| 697 |
else if( TP[i] == HPL_BLONG ) iwork[j] = 4; |
| 698 |
else if( TP[i] == HPL_BLONG_M ) iwork[j] = 5; |
| 699 |
j++; |
| 700 |
} |
| 701 |
for( i = 0; i < *NDHS; i++ ) { iwork[j] = DH[i]; j++; } |
| 702 |
|
| 703 |
if( *FSWAP == HPL_SWAP00 ) iwork[j] = 0; |
| 704 |
else if( *FSWAP == HPL_SWAP01 ) iwork[j] = 1; |
| 705 |
else if( *FSWAP == HPL_SW_MIX ) iwork[j] = 2; |
| 706 |
j++; |
| 707 |
} |
| 708 |
(void) HPL_broadcast( (void*)iwork, lwork, HPL_INT, 0, |
| 709 |
MPI_COMM_WORLD ); |
| 710 |
if( rank != 0 ) |
| 711 |
{
|
| 712 |
j = 0;
|
| 713 |
for( i = 0; i < *NS; i++ ) { N [i] = iwork[j]; j++; } |
| 714 |
for( i = 0; i < *NBS; i++ ) { NB[i] = iwork[j]; j++; } |
| 715 |
for( i = 0; i < *NPQS; i++ ) { P [i] = iwork[j]; j++; } |
| 716 |
for( i = 0; i < *NPQS; i++ ) { Q [i] = iwork[j]; j++; } |
| 717 |
|
| 718 |
for( i = 0; i < *NPFS; i++ ) |
| 719 |
{
|
| 720 |
if( iwork[j] == 0 ) PF[i] = HPL_LEFT_LOOKING; |
| 721 |
else if( iwork[j] == 1 ) PF[i] = HPL_CROUT; |
| 722 |
else if( iwork[j] == 2 ) PF[i] = HPL_RIGHT_LOOKING; |
| 723 |
j++; |
| 724 |
} |
| 725 |
for( i = 0; i < *NBMS; i++ ) { NBM[i] = iwork[j]; j++; } |
| 726 |
for( i = 0; i < *NDVS; i++ ) { NDV[i] = iwork[j]; j++; } |
| 727 |
for( i = 0; i < *NRFS; i++ ) |
| 728 |
{
|
| 729 |
if( iwork[j] == 0 ) RF[i] = HPL_LEFT_LOOKING; |
| 730 |
else if( iwork[j] == 1 ) RF[i] = HPL_CROUT; |
| 731 |
else if( iwork[j] == 2 ) RF[i] = HPL_RIGHT_LOOKING; |
| 732 |
j++; |
| 733 |
} |
| 734 |
for( i = 0; i < *NTPS; i++ ) |
| 735 |
{
|
| 736 |
if( iwork[j] == 0 ) TP[i] = HPL_1RING; |
| 737 |
else if( iwork[j] == 1 ) TP[i] = HPL_1RING_M; |
| 738 |
else if( iwork[j] == 2 ) TP[i] = HPL_2RING; |
| 739 |
else if( iwork[j] == 3 ) TP[i] = HPL_2RING_M; |
| 740 |
else if( iwork[j] == 4 ) TP[i] = HPL_BLONG; |
| 741 |
else if( iwork[j] == 5 ) TP[i] = HPL_BLONG_M; |
| 742 |
j++; |
| 743 |
} |
| 744 |
for( i = 0; i < *NDHS; i++ ) { DH[i] = iwork[j]; j++; } |
| 745 |
|
| 746 |
if( iwork[j] == 0 ) *FSWAP = HPL_SWAP00; |
| 747 |
else if( iwork[j] == 1 ) *FSWAP = HPL_SWAP01; |
| 748 |
else if( iwork[j] == 2 ) *FSWAP = HPL_SW_MIX; |
| 749 |
j++; |
| 750 |
} |
| 751 |
if( iwork ) free( iwork );
|
| 752 |
/*
|
| 753 |
* regurgitate input
|
| 754 |
*/
|
| 755 |
if( rank == 0 ) |
| 756 |
{
|
| 757 |
HPL_fprintf( TEST->outfp, "%s%s\n",
|
| 758 |
"========================================",
|
| 759 |
"========================================" );
|
| 760 |
HPL_fprintf( TEST->outfp, "%s%s\n",
|
| 761 |
"HPLinpack 2.0 -- High-Performance Linpack benchmark -- ",
|
| 762 |
" September 10, 2008" );
|
| 763 |
HPL_fprintf( TEST->outfp, "%s%s\n",
|
| 764 |
"Written by A. Petitet and R. Clint Whaley, ",
|
| 765 |
"Innovative Computing Laboratory, UTK" );
|
| 766 |
HPL_fprintf( TEST->outfp, "%s%s\n",
|
| 767 |
"Modified by Piotr Luszczek, ",
|
| 768 |
"Innovative Computing Laboratory, UTK" );
|
| 769 |
HPL_fprintf( TEST->outfp, "%s%s\n",
|
| 770 |
"Modified by Julien Langou, ",
|
| 771 |
"University of Colorado Denver");
|
| 772 |
HPL_fprintf( TEST->outfp, "%s%s\n",
|
| 773 |
"========================================",
|
| 774 |
"========================================" );
|
| 775 |
|
| 776 |
HPL_fprintf( TEST->outfp, "\n%s\n",
|
| 777 |
"An explanation of the input/output parameters follows:" );
|
| 778 |
HPL_fprintf( TEST->outfp, "%s\n",
|
| 779 |
"T/V : Wall time / encoded variant." );
|
| 780 |
HPL_fprintf( TEST->outfp, "%s\n",
|
| 781 |
"N : The order of the coefficient matrix A." );
|
| 782 |
HPL_fprintf( TEST->outfp, "%s\n",
|
| 783 |
"NB : The partitioning blocking factor." );
|
| 784 |
HPL_fprintf( TEST->outfp, "%s\n",
|
| 785 |
"P : The number of process rows." );
|
| 786 |
HPL_fprintf( TEST->outfp, "%s\n",
|
| 787 |
"Q : The number of process columns." );
|
| 788 |
HPL_fprintf( TEST->outfp, "%s\n",
|
| 789 |
"Time : Time in seconds to solve the linear system." );
|
| 790 |
HPL_fprintf( TEST->outfp, "%s\n\n",
|
| 791 |
"Gflops : Rate of execution for solving the linear system." );
|
| 792 |
HPL_fprintf( TEST->outfp, "%s\n",
|
| 793 |
"The following parameter values will be used:" );
|
| 794 |
/*
|
| 795 |
* Problem size
|
| 796 |
*/
|
| 797 |
HPL_fprintf( TEST->outfp, "\nN :" );
|
| 798 |
for( i = 0; i < Mmin( 8, *NS ); i++ ) |
| 799 |
HPL_fprintf( TEST->outfp, "%8d ", N[i] );
|
| 800 |
if( *NS > 8 ) |
| 801 |
{
|
| 802 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 803 |
for( i = 8; i < Mmin( 16, *NS ); i++ ) |
| 804 |
HPL_fprintf( TEST->outfp, "%8d ", N[i] );
|
| 805 |
if( *NS > 16 ) |
| 806 |
{
|
| 807 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 808 |
for( i = 16; i < *NS; i++ ) |
| 809 |
HPL_fprintf( TEST->outfp, "%8d ", N[i] );
|
| 810 |
} |
| 811 |
} |
| 812 |
/*
|
| 813 |
* Distribution blocking factor
|
| 814 |
*/
|
| 815 |
HPL_fprintf( TEST->outfp, "\nNB :" );
|
| 816 |
for( i = 0; i < Mmin( 8, *NBS ); i++ ) |
| 817 |
HPL_fprintf( TEST->outfp, "%8d ", NB[i] );
|
| 818 |
if( *NBS > 8 ) |
| 819 |
{
|
| 820 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 821 |
for( i = 8; i < Mmin( 16, *NBS ); i++ ) |
| 822 |
HPL_fprintf( TEST->outfp, "%8d ", NB[i] );
|
| 823 |
if( *NBS > 16 ) |
| 824 |
{
|
| 825 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 826 |
for( i = 16; i < *NBS; i++ ) |
| 827 |
HPL_fprintf( TEST->outfp, "%8d ", NB[i] );
|
| 828 |
} |
| 829 |
} |
| 830 |
/*
|
| 831 |
* Process mapping
|
| 832 |
*/
|
| 833 |
HPL_fprintf( TEST->outfp, "\nPMAP :" );
|
| 834 |
if( *PMAPPIN == HPL_ROW_MAJOR )
|
| 835 |
HPL_fprintf( TEST->outfp, " Row-major process mapping" );
|
| 836 |
else if( *PMAPPIN == HPL_COLUMN_MAJOR ) |
| 837 |
HPL_fprintf( TEST->outfp, " Column-major process mapping" );
|
| 838 |
/*
|
| 839 |
* Process grid
|
| 840 |
*/
|
| 841 |
HPL_fprintf( TEST->outfp, "\nP :" );
|
| 842 |
for( i = 0; i < Mmin( 8, *NPQS ); i++ ) |
| 843 |
HPL_fprintf( TEST->outfp, "%8d ", P[i] );
|
| 844 |
if( *NPQS > 8 ) |
| 845 |
{
|
| 846 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 847 |
for( i = 8; i < Mmin( 16, *NPQS ); i++ ) |
| 848 |
HPL_fprintf( TEST->outfp, "%8d ", P[i] );
|
| 849 |
if( *NPQS > 16 ) |
| 850 |
{
|
| 851 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 852 |
for( i = 16; i < *NPQS; i++ ) |
| 853 |
HPL_fprintf( TEST->outfp, "%8d ", P[i] );
|
| 854 |
} |
| 855 |
} |
| 856 |
HPL_fprintf( TEST->outfp, "\nQ :" );
|
| 857 |
for( i = 0; i < Mmin( 8, *NPQS ); i++ ) |
| 858 |
HPL_fprintf( TEST->outfp, "%8d ", Q[i] );
|
| 859 |
if( *NPQS > 8 ) |
| 860 |
{
|
| 861 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 862 |
for( i = 8; i < Mmin( 16, *NPQS ); i++ ) |
| 863 |
HPL_fprintf( TEST->outfp, "%8d ", Q[i] );
|
| 864 |
if( *NPQS > 16 ) |
| 865 |
{
|
| 866 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 867 |
for( i = 16; i < *NPQS; i++ ) |
| 868 |
HPL_fprintf( TEST->outfp, "%8d ", Q[i] );
|
| 869 |
} |
| 870 |
} |
| 871 |
/*
|
| 872 |
* Panel Factorization
|
| 873 |
*/
|
| 874 |
HPL_fprintf( TEST->outfp, "\nPFACT :" );
|
| 875 |
for( i = 0; i < Mmin( 8, *NPFS ); i++ ) |
| 876 |
{
|
| 877 |
if( PF[i] == HPL_LEFT_LOOKING )
|
| 878 |
HPL_fprintf( TEST->outfp, " Left " );
|
| 879 |
else if( PF[i] == HPL_CROUT ) |
| 880 |
HPL_fprintf( TEST->outfp, " Crout " );
|
| 881 |
else if( PF[i] == HPL_RIGHT_LOOKING ) |
| 882 |
HPL_fprintf( TEST->outfp, " Right " );
|
| 883 |
} |
| 884 |
if( *NPFS > 8 ) |
| 885 |
{
|
| 886 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 887 |
for( i = 8; i < Mmin( 16, *NPFS ); i++ ) |
| 888 |
{
|
| 889 |
if( PF[i] == HPL_LEFT_LOOKING )
|
| 890 |
HPL_fprintf( TEST->outfp, " Left " );
|
| 891 |
else if( PF[i] == HPL_CROUT ) |
| 892 |
HPL_fprintf( TEST->outfp, " Crout " );
|
| 893 |
else if( PF[i] == HPL_RIGHT_LOOKING ) |
| 894 |
HPL_fprintf( TEST->outfp, " Right " );
|
| 895 |
} |
| 896 |
if( *NPFS > 16 ) |
| 897 |
{
|
| 898 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 899 |
for( i = 16; i < *NPFS; i++ ) |
| 900 |
{
|
| 901 |
if( PF[i] == HPL_LEFT_LOOKING )
|
| 902 |
HPL_fprintf( TEST->outfp, " Left " );
|
| 903 |
else if( PF[i] == HPL_CROUT ) |
| 904 |
HPL_fprintf( TEST->outfp, " Crout " );
|
| 905 |
else if( PF[i] == HPL_RIGHT_LOOKING ) |
| 906 |
HPL_fprintf( TEST->outfp, " Right " );
|
| 907 |
} |
| 908 |
} |
| 909 |
} |
| 910 |
/*
|
| 911 |
* Recursive stopping criterium
|
| 912 |
*/
|
| 913 |
HPL_fprintf( TEST->outfp, "\nNBMIN :" );
|
| 914 |
for( i = 0; i < Mmin( 8, *NBMS ); i++ ) |
| 915 |
HPL_fprintf( TEST->outfp, "%8d ", NBM[i] );
|
| 916 |
if( *NBMS > 8 ) |
| 917 |
{
|
| 918 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 919 |
for( i = 8; i < Mmin( 16, *NBMS ); i++ ) |
| 920 |
HPL_fprintf( TEST->outfp, "%8d ", NBM[i] );
|
| 921 |
if( *NBMS > 16 ) |
| 922 |
{
|
| 923 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 924 |
for( i = 16; i < *NBMS; i++ ) |
| 925 |
HPL_fprintf( TEST->outfp, "%8d ", NBM[i] );
|
| 926 |
} |
| 927 |
} |
| 928 |
/*
|
| 929 |
* Number of panels in recursion
|
| 930 |
*/
|
| 931 |
HPL_fprintf( TEST->outfp, "\nNDIV :" );
|
| 932 |
for( i = 0; i < Mmin( 8, *NDVS ); i++ ) |
| 933 |
HPL_fprintf( TEST->outfp, "%8d ", NDV[i] );
|
| 934 |
if( *NDVS > 8 ) |
| 935 |
{
|
| 936 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 937 |
for( i = 8; i < Mmin( 16, *NDVS ); i++ ) |
| 938 |
HPL_fprintf( TEST->outfp, "%8d ", NDV[i] );
|
| 939 |
if( *NDVS > 16 ) |
| 940 |
{
|
| 941 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 942 |
for( i = 16; i < *NDVS; i++ ) |
| 943 |
HPL_fprintf( TEST->outfp, "%8d ", NDV[i] );
|
| 944 |
} |
| 945 |
} |
| 946 |
/*
|
| 947 |
* Recursive Factorization
|
| 948 |
*/
|
| 949 |
HPL_fprintf( TEST->outfp, "\nRFACT :" );
|
| 950 |
for( i = 0; i < Mmin( 8, *NRFS ); i++ ) |
| 951 |
{
|
| 952 |
if( RF[i] == HPL_LEFT_LOOKING )
|
| 953 |
HPL_fprintf( TEST->outfp, " Left " );
|
| 954 |
else if( RF[i] == HPL_CROUT ) |
| 955 |
HPL_fprintf( TEST->outfp, " Crout " );
|
| 956 |
else if( RF[i] == HPL_RIGHT_LOOKING ) |
| 957 |
HPL_fprintf( TEST->outfp, " Right " );
|
| 958 |
} |
| 959 |
if( *NRFS > 8 ) |
| 960 |
{
|
| 961 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 962 |
for( i = 8; i < Mmin( 16, *NRFS ); i++ ) |
| 963 |
{
|
| 964 |
if( RF[i] == HPL_LEFT_LOOKING )
|
| 965 |
HPL_fprintf( TEST->outfp, " Left " );
|
| 966 |
else if( RF[i] == HPL_CROUT ) |
| 967 |
HPL_fprintf( TEST->outfp, " Crout " );
|
| 968 |
else if( RF[i] == HPL_RIGHT_LOOKING ) |
| 969 |
HPL_fprintf( TEST->outfp, " Right " );
|
| 970 |
} |
| 971 |
if( *NRFS > 16 ) |
| 972 |
{
|
| 973 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 974 |
for( i = 16; i < *NRFS; i++ ) |
| 975 |
{
|
| 976 |
if( RF[i] == HPL_LEFT_LOOKING )
|
| 977 |
HPL_fprintf( TEST->outfp, " Left " );
|
| 978 |
else if( RF[i] == HPL_CROUT ) |
| 979 |
HPL_fprintf( TEST->outfp, " Crout " );
|
| 980 |
else if( RF[i] == HPL_RIGHT_LOOKING ) |
| 981 |
HPL_fprintf( TEST->outfp, " Right " );
|
| 982 |
} |
| 983 |
} |
| 984 |
} |
| 985 |
/*
|
| 986 |
* Broadcast topology
|
| 987 |
*/
|
| 988 |
HPL_fprintf( TEST->outfp, "\nBCAST :" );
|
| 989 |
for( i = 0; i < Mmin( 8, *NTPS ); i++ ) |
| 990 |
{
|
| 991 |
if( TP[i] == HPL_1RING )
|
| 992 |
HPL_fprintf( TEST->outfp, " 1ring " );
|
| 993 |
else if( TP[i] == HPL_1RING_M ) |
| 994 |
HPL_fprintf( TEST->outfp, " 1ringM " );
|
| 995 |
else if( TP[i] == HPL_2RING ) |
| 996 |
HPL_fprintf( TEST->outfp, " 2ring " );
|
| 997 |
else if( TP[i] == HPL_2RING_M ) |
| 998 |
HPL_fprintf( TEST->outfp, " 2ringM " );
|
| 999 |
else if( TP[i] == HPL_BLONG ) |
| 1000 |
HPL_fprintf( TEST->outfp, " Blong " );
|
| 1001 |
else if( TP[i] == HPL_BLONG_M ) |
| 1002 |
HPL_fprintf( TEST->outfp, " BlongM " );
|
| 1003 |
} |
| 1004 |
if( *NTPS > 8 ) |
| 1005 |
{
|
| 1006 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 1007 |
for( i = 8; i < Mmin( 16, *NTPS ); i++ ) |
| 1008 |
{
|
| 1009 |
if( TP[i] == HPL_1RING )
|
| 1010 |
HPL_fprintf( TEST->outfp, " 1ring " );
|
| 1011 |
else if( TP[i] == HPL_1RING_M ) |
| 1012 |
HPL_fprintf( TEST->outfp, " 1ringM " );
|
| 1013 |
else if( TP[i] == HPL_2RING ) |
| 1014 |
HPL_fprintf( TEST->outfp, " 2ring " );
|
| 1015 |
else if( TP[i] == HPL_2RING_M ) |
| 1016 |
HPL_fprintf( TEST->outfp, " 2ringM " );
|
| 1017 |
else if( TP[i] == HPL_BLONG ) |
| 1018 |
HPL_fprintf( TEST->outfp, " Blong " );
|
| 1019 |
else if( TP[i] == HPL_BLONG_M ) |
| 1020 |
HPL_fprintf( TEST->outfp, " BlongM " );
|
| 1021 |
} |
| 1022 |
if( *NTPS > 16 ) |
| 1023 |
{
|
| 1024 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 1025 |
for( i = 16; i < *NTPS; i++ ) |
| 1026 |
{
|
| 1027 |
if( TP[i] == HPL_1RING )
|
| 1028 |
HPL_fprintf( TEST->outfp, " 1ring " );
|
| 1029 |
else if( TP[i] == HPL_1RING_M ) |
| 1030 |
HPL_fprintf( TEST->outfp, " 1ringM " );
|
| 1031 |
else if( TP[i] == HPL_2RING ) |
| 1032 |
HPL_fprintf( TEST->outfp, " 2ring " );
|
| 1033 |
else if( TP[i] == HPL_2RING_M ) |
| 1034 |
HPL_fprintf( TEST->outfp, " 2ringM " );
|
| 1035 |
else if( TP[i] == HPL_BLONG ) |
| 1036 |
HPL_fprintf( TEST->outfp, " Blong " );
|
| 1037 |
else if( TP[i] == HPL_BLONG_M ) |
| 1038 |
HPL_fprintf( TEST->outfp, " BlongM " );
|
| 1039 |
} |
| 1040 |
} |
| 1041 |
} |
| 1042 |
/*
|
| 1043 |
* Lookahead depths
|
| 1044 |
*/
|
| 1045 |
HPL_fprintf( TEST->outfp, "\nDEPTH :" );
|
| 1046 |
for( i = 0; i < Mmin( 8, *NDHS ); i++ ) |
| 1047 |
HPL_fprintf( TEST->outfp, "%8d ", DH[i] );
|
| 1048 |
if( *NDHS > 8 ) |
| 1049 |
{
|
| 1050 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 1051 |
for( i = 8; i < Mmin( 16, *NDHS ); i++ ) |
| 1052 |
HPL_fprintf( TEST->outfp, "%8d ", DH[i] );
|
| 1053 |
if( *NDHS > 16 ) |
| 1054 |
{
|
| 1055 |
HPL_fprintf( TEST->outfp, "\n " );
|
| 1056 |
for( i = 16; i < *NDHS; i++ ) |
| 1057 |
HPL_fprintf( TEST->outfp, "%8d ", DH[i] );
|
| 1058 |
} |
| 1059 |
} |
| 1060 |
/*
|
| 1061 |
* Swapping algorithm
|
| 1062 |
*/
|
| 1063 |
HPL_fprintf( TEST->outfp, "\nSWAP :" );
|
| 1064 |
if( *FSWAP == HPL_SWAP00 )
|
| 1065 |
HPL_fprintf( TEST->outfp, " Binary-exchange" );
|
| 1066 |
else if( *FSWAP == HPL_SWAP01 ) |
| 1067 |
HPL_fprintf( TEST->outfp, " Spread-roll (long)" );
|
| 1068 |
else if( *FSWAP == HPL_SW_MIX ) |
| 1069 |
HPL_fprintf( TEST->outfp, " Mix (threshold = %d)", *TSWAP );
|
| 1070 |
/*
|
| 1071 |
* L1 storage form
|
| 1072 |
*/
|
| 1073 |
HPL_fprintf( TEST->outfp, "\nL1 :" );
|
| 1074 |
if( *L1NOTRAN != 0 ) |
| 1075 |
HPL_fprintf( TEST->outfp, " no-transposed form" );
|
| 1076 |
else
|
| 1077 |
HPL_fprintf( TEST->outfp, " transposed form" );
|
| 1078 |
/*
|
| 1079 |
* U storage form
|
| 1080 |
*/
|
| 1081 |
HPL_fprintf( TEST->outfp, "\nU :" );
|
| 1082 |
if( *UNOTRAN != 0 ) |
| 1083 |
HPL_fprintf( TEST->outfp, " no-transposed form" );
|
| 1084 |
else
|
| 1085 |
HPL_fprintf( TEST->outfp, " transposed form" );
|
| 1086 |
/*
|
| 1087 |
* Equilibration
|
| 1088 |
*/
|
| 1089 |
HPL_fprintf( TEST->outfp, "\nEQUIL :" );
|
| 1090 |
if( *EQUIL != 0 ) |
| 1091 |
HPL_fprintf( TEST->outfp, " yes" );
|
| 1092 |
else
|
| 1093 |
HPL_fprintf( TEST->outfp, " no" );
|
| 1094 |
/*
|
| 1095 |
* Alignment
|
| 1096 |
*/
|
| 1097 |
HPL_fprintf( TEST->outfp, "\nALIGN : %d double precision words",
|
| 1098 |
*ALIGN ); |
| 1099 |
|
| 1100 |
HPL_fprintf( TEST->outfp, "\n\n" );
|
| 1101 |
/*
|
| 1102 |
* For testing only
|
| 1103 |
*/
|
| 1104 |
if( TEST->thrsh > HPL_rzero )
|
| 1105 |
{
|
| 1106 |
HPL_fprintf( TEST->outfp, "%s%s\n\n",
|
| 1107 |
"----------------------------------------",
|
| 1108 |
"----------------------------------------" );
|
| 1109 |
HPL_fprintf( TEST->outfp, "%s\n",
|
| 1110 |
"- The matrix A is randomly generated for each test." );
|
| 1111 |
HPL_fprintf( TEST->outfp, "%s\n",
|
| 1112 |
"- The following scaled residual check will be computed:" );
|
| 1113 |
HPL_fprintf( TEST->outfp, "%s\n",
|
| 1114 |
" ||Ax-b||_oo / ( eps * ( || x ||_oo * || A ||_oo + || b ||_oo ) * N )" );
|
| 1115 |
HPL_fprintf( TEST->outfp, "%s %21.6e\n",
|
| 1116 |
"- The relative machine precision (eps) is taken to be ",
|
| 1117 |
TEST->epsil ); |
| 1118 |
HPL_fprintf( TEST->outfp, "%s %11.1f\n\n",
|
| 1119 |
"- Computational tests pass if scaled residuals are less than ",
|
| 1120 |
TEST->thrsh ); |
| 1121 |
} |
| 1122 |
} |
| 1123 |
/*
|
| 1124 |
* End of HPL_pdinfo
|
| 1125 |
*/
|
| 1126 |
} |