root / src / pgesv / HPL_pdupdateNT.c @ 9
Historique | Voir | Annoter | Télécharger (15,15 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_pdupdateNT
|
| 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_pdupdateNT
|
| 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_pdupdateNT broadcast - forward the panel PBCST and simultaneously
|
| 74 | 1 | equemene | * applies the row interchanges and updates part of the trailing (using
|
| 75 | 1 | equemene | * the panel PANEL) submatrix.
|
| 76 | 1 | equemene | *
|
| 77 | 1 | equemene | * Arguments
|
| 78 | 1 | equemene | * =========
|
| 79 | 1 | equemene | *
|
| 80 | 1 | equemene | * PBCST (local input/output) HPL_T_panel *
|
| 81 | 1 | equemene | * On entry, PBCST points to the data structure containing the
|
| 82 | 1 | equemene | * panel (to be broadcast) information.
|
| 83 | 1 | equemene | *
|
| 84 | 1 | equemene | * IFLAG (local output) int *
|
| 85 | 1 | equemene | * On exit, IFLAG indicates whether or not the broadcast has
|
| 86 | 1 | equemene | * been completed when PBCST is not NULL on entry. In that case,
|
| 87 | 1 | equemene | * IFLAG is left unchanged.
|
| 88 | 1 | equemene | *
|
| 89 | 1 | equemene | * PANEL (local input/output) HPL_T_panel *
|
| 90 | 1 | equemene | * On entry, PANEL points to the data structure containing the
|
| 91 | 1 | equemene | * panel (to be updated) information.
|
| 92 | 1 | equemene | *
|
| 93 | 1 | equemene | * NN (local input) const int
|
| 94 | 1 | equemene | * On entry, NN specifies the local number of columns of the
|
| 95 | 1 | equemene | * trailing submatrix to be updated starting at the current
|
| 96 | 1 | equemene | * position. NN must be at least zero.
|
| 97 | 1 | equemene | *
|
| 98 | 1 | equemene | * ---------------------------------------------------------------------
|
| 99 | 1 | equemene | */
|
| 100 | 1 | equemene | /*
|
| 101 | 1 | equemene | * .. Local Variables ..
|
| 102 | 1 | equemene | */
|
| 103 | 1 | equemene | double * Aptr, * L1ptr, * L2ptr, * Uptr, * dpiv;
|
| 104 | 1 | equemene | int * ipiv;
|
| 105 | 1 | equemene | #ifdef HPL_CALL_VSIPL
|
| 106 | 1 | equemene | vsip_mview_d * Av0, * Av1, * Lv0, * Lv1, * Uv0, * Uv1; |
| 107 | 1 | equemene | #endif
|
| 108 | 1 | equemene | int curr, i, iroff, jb, lda, ldl2, mp, n, nb,
|
| 109 | 1 | equemene | nq0, nn, test; |
| 110 | 1 | equemene | static int tswap = 0; |
| 111 | 1 | equemene | static HPL_T_SWAP fswap = HPL_NO_SWP;
|
| 112 | 1 | equemene | #define LDU n
|
| 113 | 1 | equemene | /* ..
|
| 114 | 1 | equemene | * .. Executable Statements ..
|
| 115 | 1 | equemene | */
|
| 116 | 1 | equemene | #ifdef HPL_DETAILED_TIMING
|
| 117 | 1 | equemene | HPL_ptimer( HPL_TIMING_UPDATE ); |
| 118 | 1 | equemene | #endif
|
| 119 | 1 | equemene | nb = PANEL->nb; jb = PANEL->jb; n = PANEL->nq; lda = PANEL->lda; |
| 120 | 1 | equemene | if( NN >= 0 ) n = Mmin( NN, n ); |
| 121 | 1 | equemene | /*
|
| 122 | 1 | equemene | * There is nothing to update, enforce the panel broadcast.
|
| 123 | 1 | equemene | */
|
| 124 | 1 | equemene | if( ( n <= 0 ) || ( jb <= 0 ) ) |
| 125 | 1 | equemene | {
|
| 126 | 1 | equemene | if( PBCST != NULL ) |
| 127 | 1 | equemene | {
|
| 128 | 1 | equemene | do { (void) HPL_bcast( PBCST, IFLAG ); } |
| 129 | 1 | equemene | while( *IFLAG != HPL_SUCCESS );
|
| 130 | 1 | equemene | } |
| 131 | 1 | equemene | #ifdef HPL_DETAILED_TIMING
|
| 132 | 1 | equemene | HPL_ptimer( HPL_TIMING_UPDATE ); |
| 133 | 1 | equemene | #endif
|
| 134 | 1 | equemene | return;
|
| 135 | 1 | equemene | } |
| 136 | 1 | equemene | /*
|
| 137 | 1 | equemene | * Enable/disable the column panel probing mechanism
|
| 138 | 1 | equemene | */
|
| 139 | 1 | equemene | (void) HPL_bcast( PBCST, &test );
|
| 140 | 1 | equemene | /*
|
| 141 | 1 | equemene | * 1 x Q case
|
| 142 | 1 | equemene | */
|
| 143 | 1 | equemene | if( PANEL->grid->nprow == 1 ) |
| 144 | 1 | equemene | {
|
| 145 | 1 | equemene | Aptr = PANEL->A; L2ptr = PANEL->L2; L1ptr = PANEL->L1; |
| 146 | 1 | equemene | ldl2 = PANEL->ldl2; dpiv = PANEL->DPIV; ipiv = PANEL->IWORK; |
| 147 | 1 | equemene | mp = PANEL->mp - jb; iroff = PANEL->ii; nq0 = 0;
|
| 148 | 1 | equemene | #ifdef HPL_CALL_VSIPL
|
| 149 | 1 | equemene | /*
|
| 150 | 1 | equemene | * Admit the blocks
|
| 151 | 1 | equemene | */
|
| 152 | 1 | equemene | (void) vsip_blockadmit_d( PANEL->Ablock, VSIP_TRUE );
|
| 153 | 1 | equemene | (void) vsip_blockadmit_d( PANEL->L2block, VSIP_TRUE );
|
| 154 | 1 | equemene | /*
|
| 155 | 1 | equemene | * Create the matrix views
|
| 156 | 1 | equemene | */
|
| 157 | 1 | equemene | Av0 = vsip_mbind_d( PANEL->Ablock, 0, 1, lda, lda, PANEL->pmat->nq ); |
| 158 | 1 | equemene | Lv0 = vsip_mbind_d( PANEL->L2block, 0, 1, ldl2, ldl2, jb ); |
| 159 | 1 | equemene | /*
|
| 160 | 1 | equemene | * Create the matrix subviews
|
| 161 | 1 | equemene | */
|
| 162 | 1 | equemene | Lv1 = vsip_msubview_d( Lv0, 0, 0, mp, jb ); |
| 163 | 1 | equemene | #endif
|
| 164 | 1 | equemene | for( i = 0; i < jb; i++ ) { ipiv[i] = (int)(dpiv[i]) - iroff; } |
| 165 | 1 | equemene | /*
|
| 166 | 1 | equemene | * So far we have not updated anything - test availability of the panel
|
| 167 | 1 | equemene | * to be forwarded - If detected forward it and finish the update in one
|
| 168 | 1 | equemene | * step.
|
| 169 | 1 | equemene | */
|
| 170 | 1 | equemene | while ( test == HPL_KEEP_TESTING )
|
| 171 | 1 | equemene | {
|
| 172 | 1 | equemene | nn = n - nq0; nn = Mmin( nb, nn ); |
| 173 | 1 | equemene | /*
|
| 174 | 1 | equemene | * Update nb columns at a time
|
| 175 | 1 | equemene | */
|
| 176 | 1 | equemene | #ifdef HPL_DETAILED_TIMING
|
| 177 | 1 | equemene | HPL_ptimer( HPL_TIMING_LASWP ); |
| 178 | 1 | equemene | HPL_dlaswp00N( jb, nn, Aptr, lda, ipiv ); |
| 179 | 1 | equemene | HPL_ptimer( HPL_TIMING_LASWP ); |
| 180 | 1 | equemene | #else
|
| 181 | 1 | equemene | HPL_dlaswp00N( jb, nn, Aptr, lda, ipiv ); |
| 182 | 1 | equemene | #endif
|
| 183 | 1 | equemene | HPL_dtrsm( HplColumnMajor, HplLeft, HplLower, HplNoTrans, |
| 184 | 1 | equemene | HplUnit, jb, nn, HPL_rone, L1ptr, jb, Aptr, lda ); |
| 185 | 1 | equemene | #ifdef HPL_CALL_VSIPL
|
| 186 | 1 | equemene | /*
|
| 187 | 1 | equemene | * Create the matrix subviews
|
| 188 | 1 | equemene | */
|
| 189 | 1 | equemene | Uv1 = vsip_msubview_d( Av0, PANEL->ii, PANEL->jj+nq0, jb, nn ); |
| 190 | 1 | equemene | Av1 = vsip_msubview_d( Av0, PANEL->ii+jb, PANEL->jj+nq0, mp, nn ); |
| 191 | 1 | equemene | |
| 192 | 1 | equemene | vsip_gemp_d( -HPL_rone, Lv1, VSIP_MAT_NTRANS, Uv1, VSIP_MAT_NTRANS, |
| 193 | 1 | equemene | HPL_rone, Av1 ); |
| 194 | 1 | equemene | /*
|
| 195 | 1 | equemene | * Destroy the matrix subviews
|
| 196 | 1 | equemene | */
|
| 197 | 1 | equemene | (void) vsip_mdestroy_d( Av1 );
|
| 198 | 1 | equemene | (void) vsip_mdestroy_d( Uv1 );
|
| 199 | 1 | equemene | #else
|
| 200 | 1 | equemene | HPL_dgemm( HplColumnMajor, HplNoTrans, HplNoTrans, mp, nn, |
| 201 | 1 | equemene | jb, -HPL_rone, L2ptr, ldl2, Aptr, lda, HPL_rone, |
| 202 | 1 | equemene | Mptr( Aptr, jb, 0, lda ), lda );
|
| 203 | 1 | equemene | #endif
|
| 204 | 1 | equemene | Aptr = Mptr( Aptr, 0, nn, lda ); nq0 += nn;
|
| 205 | 1 | equemene | |
| 206 | 1 | equemene | (void) HPL_bcast( PBCST, &test );
|
| 207 | 1 | equemene | } |
| 208 | 1 | equemene | /*
|
| 209 | 1 | equemene | * The panel has been forwarded at that point, finish the update
|
| 210 | 1 | equemene | */
|
| 211 | 1 | equemene | if( ( nn = n - nq0 ) > 0 ) |
| 212 | 1 | equemene | {
|
| 213 | 1 | equemene | #ifdef HPL_DETAILED_TIMING
|
| 214 | 1 | equemene | HPL_ptimer( HPL_TIMING_LASWP ); |
| 215 | 1 | equemene | HPL_dlaswp00N( jb, nn, Aptr, lda, ipiv ); |
| 216 | 1 | equemene | HPL_ptimer( HPL_TIMING_LASWP ); |
| 217 | 1 | equemene | #else
|
| 218 | 1 | equemene | HPL_dlaswp00N( jb, nn, Aptr, lda, ipiv ); |
| 219 | 1 | equemene | #endif
|
| 220 | 1 | equemene | HPL_dtrsm( HplColumnMajor, HplLeft, HplLower, HplNoTrans, |
| 221 | 1 | equemene | HplUnit, jb, nn, HPL_rone, L1ptr, jb, Aptr, lda ); |
| 222 | 1 | equemene | #ifdef HPL_CALL_VSIPL
|
| 223 | 1 | equemene | /*
|
| 224 | 1 | equemene | * Create the matrix subviews
|
| 225 | 1 | equemene | */
|
| 226 | 1 | equemene | Uv1 = vsip_msubview_d( Av0, PANEL->ii, PANEL->jj+nq0, jb, nn ); |
| 227 | 1 | equemene | Av1 = vsip_msubview_d( Av0, PANEL->ii+jb, PANEL->jj+nq0, mp, nn ); |
| 228 | 1 | equemene | |
| 229 | 1 | equemene | vsip_gemp_d( -HPL_rone, Lv1, VSIP_MAT_NTRANS, Uv1, VSIP_MAT_NTRANS, |
| 230 | 1 | equemene | HPL_rone, Av1 ); |
| 231 | 1 | equemene | /*
|
| 232 | 1 | equemene | * Destroy the matrix subviews
|
| 233 | 1 | equemene | */
|
| 234 | 1 | equemene | (void) vsip_mdestroy_d( Av1 );
|
| 235 | 1 | equemene | (void) vsip_mdestroy_d( Uv1 );
|
| 236 | 1 | equemene | #else
|
| 237 | 1 | equemene | HPL_dgemm( HplColumnMajor, HplNoTrans, HplNoTrans, mp, nn, |
| 238 | 1 | equemene | jb, -HPL_rone, L2ptr, ldl2, Aptr, lda, HPL_rone, |
| 239 | 1 | equemene | Mptr( Aptr, jb, 0, lda ), lda );
|
| 240 | 1 | equemene | #endif
|
| 241 | 1 | equemene | } |
| 242 | 1 | equemene | #ifdef HPL_CALL_VSIPL
|
| 243 | 1 | equemene | /*
|
| 244 | 1 | equemene | * Destroy the matrix subviews
|
| 245 | 1 | equemene | */
|
| 246 | 1 | equemene | (void) vsip_mdestroy_d( Lv1 );
|
| 247 | 1 | equemene | /*
|
| 248 | 1 | equemene | * Release the blocks
|
| 249 | 1 | equemene | */
|
| 250 | 1 | equemene | (void) vsip_blockrelease_d( vsip_mgetblock_d( Lv0 ), VSIP_TRUE );
|
| 251 | 1 | equemene | (void) vsip_blockrelease_d( vsip_mgetblock_d( Av0 ), VSIP_TRUE );
|
| 252 | 1 | equemene | /*
|
| 253 | 1 | equemene | * Destroy the matrix views
|
| 254 | 1 | equemene | */
|
| 255 | 1 | equemene | (void) vsip_mdestroy_d( Lv0 );
|
| 256 | 1 | equemene | (void) vsip_mdestroy_d( Av0 );
|
| 257 | 1 | equemene | #endif
|
| 258 | 1 | equemene | } |
| 259 | 1 | equemene | else /* nprow > 1 ... */ |
| 260 | 1 | equemene | {
|
| 261 | 1 | equemene | /*
|
| 262 | 1 | equemene | * Selection of the swapping algorithm - swap:broadcast U.
|
| 263 | 1 | equemene | */
|
| 264 | 1 | equemene | if( fswap == HPL_NO_SWP )
|
| 265 | 1 | equemene | { fswap = PANEL->algo->fswap; tswap = PANEL->algo->fsthr; }
|
| 266 | 1 | equemene | |
| 267 | 1 | equemene | if( ( fswap == HPL_SWAP01 ) ||
|
| 268 | 1 | equemene | ( ( fswap == HPL_SW_MIX ) && ( n > tswap ) ) ) |
| 269 | 1 | equemene | { HPL_pdlaswp01T( PBCST, &test, PANEL, n ); }
|
| 270 | 1 | equemene | else
|
| 271 | 1 | equemene | { HPL_pdlaswp00T( PBCST, &test, PANEL, n ); }
|
| 272 | 1 | equemene | /*
|
| 273 | 1 | equemene | * Compute redundantly row block of U and update trailing submatrix
|
| 274 | 1 | equemene | */
|
| 275 | 1 | equemene | nq0 = 0; curr = ( PANEL->grid->myrow == PANEL->prow ? 1 : 0 ); |
| 276 | 1 | equemene | Aptr = PANEL->A; L2ptr = PANEL->L2; L1ptr = PANEL->L1; |
| 277 | 1 | equemene | Uptr = PANEL->U; ldl2 = PANEL->ldl2; |
| 278 | 1 | equemene | mp = PANEL->mp - ( curr != 0 ? jb : 0 ); |
| 279 | 1 | equemene | #ifdef HPL_CALL_VSIPL
|
| 280 | 1 | equemene | /*
|
| 281 | 1 | equemene | * Admit the blocks
|
| 282 | 1 | equemene | */
|
| 283 | 1 | equemene | (void) vsip_blockadmit_d( PANEL->Ablock, VSIP_TRUE );
|
| 284 | 1 | equemene | (void) vsip_blockadmit_d( PANEL->L2block, VSIP_TRUE );
|
| 285 | 1 | equemene | (void) vsip_blockadmit_d( PANEL->Ublock, VSIP_TRUE );
|
| 286 | 1 | equemene | /*
|
| 287 | 1 | equemene | * Create the matrix views
|
| 288 | 1 | equemene | */
|
| 289 | 1 | equemene | Av0 = vsip_mbind_d( PANEL->Ablock, 0, 1, lda, lda, PANEL->pmat->nq ); |
| 290 | 1 | equemene | Lv0 = vsip_mbind_d( PANEL->L2block, 0, 1, ldl2, ldl2, jb ); |
| 291 | 1 | equemene | Uv0 = vsip_mbind_d( PANEL->Ublock, 0, 1, LDU, LDU, jb ); |
| 292 | 1 | equemene | /*
|
| 293 | 1 | equemene | * Create the matrix subviews
|
| 294 | 1 | equemene | */
|
| 295 | 1 | equemene | Lv1 = vsip_msubview_d( Lv0, 0, 0, mp, jb ); |
| 296 | 1 | equemene | #endif
|
| 297 | 1 | equemene | /*
|
| 298 | 1 | equemene | * Broadcast has not occured yet, spliting the computational part
|
| 299 | 1 | equemene | */
|
| 300 | 1 | equemene | while ( test == HPL_KEEP_TESTING )
|
| 301 | 1 | equemene | {
|
| 302 | 1 | equemene | nn = n - nq0; nn = Mmin( nb, nn ); |
| 303 | 1 | equemene | |
| 304 | 1 | equemene | HPL_dtrsm( HplColumnMajor, HplRight, HplLower, HplTrans, |
| 305 | 1 | equemene | HplUnit, nn, jb, HPL_rone, L1ptr, jb, Uptr, LDU ); |
| 306 | 1 | equemene | |
| 307 | 1 | equemene | if( curr != 0 ) |
| 308 | 1 | equemene | {
|
| 309 | 1 | equemene | #ifdef HPL_CALL_VSIPL
|
| 310 | 1 | equemene | /*
|
| 311 | 1 | equemene | * Create the matrix subviews
|
| 312 | 1 | equemene | */
|
| 313 | 1 | equemene | Uv1 = vsip_msubview_d( Uv0, nq0, 0, nn, jb );
|
| 314 | 1 | equemene | Av1 = vsip_msubview_d( Av0, PANEL->ii+jb, PANEL->jj+nq0, mp, nn ); |
| 315 | 1 | equemene | |
| 316 | 1 | equemene | vsip_gemp_d( -HPL_rone, Lv1, VSIP_MAT_NTRANS, Uv1, VSIP_MAT_TRANS, |
| 317 | 1 | equemene | HPL_rone, Av1 ); |
| 318 | 1 | equemene | /*
|
| 319 | 1 | equemene | * Destroy the matrix subviews
|
| 320 | 1 | equemene | */
|
| 321 | 1 | equemene | (void) vsip_mdestroy_d( Av1 );
|
| 322 | 1 | equemene | (void) vsip_mdestroy_d( Uv1 );
|
| 323 | 1 | equemene | #else
|
| 324 | 1 | equemene | HPL_dgemm( HplColumnMajor, HplNoTrans, HplTrans, mp, nn, |
| 325 | 1 | equemene | jb, -HPL_rone, L2ptr, ldl2, Uptr, LDU, HPL_rone, |
| 326 | 1 | equemene | Mptr( Aptr, jb, 0, lda ), lda );
|
| 327 | 1 | equemene | #endif
|
| 328 | 1 | equemene | HPL_dlatcpy( jb, nn, Uptr, LDU, Aptr, lda ); |
| 329 | 1 | equemene | } |
| 330 | 1 | equemene | else
|
| 331 | 1 | equemene | {
|
| 332 | 1 | equemene | #ifdef HPL_CALL_VSIPL
|
| 333 | 1 | equemene | /*
|
| 334 | 1 | equemene | * Create the matrix subviews
|
| 335 | 1 | equemene | */
|
| 336 | 1 | equemene | Uv1 = vsip_msubview_d( Uv0, nq0, 0, nn, jb );
|
| 337 | 1 | equemene | Av1 = vsip_msubview_d( Av0, PANEL->ii, PANEL->jj+nq0, mp, nn ); |
| 338 | 1 | equemene | |
| 339 | 1 | equemene | vsip_gemp_d( -HPL_rone, Lv1, VSIP_MAT_NTRANS, Uv1, VSIP_MAT_TRANS, |
| 340 | 1 | equemene | HPL_rone, Av1 ); |
| 341 | 1 | equemene | /*
|
| 342 | 1 | equemene | * Destroy the matrix subviews
|
| 343 | 1 | equemene | */
|
| 344 | 1 | equemene | (void) vsip_mdestroy_d( Av1 );
|
| 345 | 1 | equemene | (void) vsip_mdestroy_d( Uv1 );
|
| 346 | 1 | equemene | #else
|
| 347 | 1 | equemene | HPL_dgemm( HplColumnMajor, HplNoTrans, HplTrans, mp, nn, |
| 348 | 1 | equemene | jb, -HPL_rone, L2ptr, ldl2, Uptr, LDU, HPL_rone, |
| 349 | 1 | equemene | Aptr, lda ); |
| 350 | 1 | equemene | #endif
|
| 351 | 1 | equemene | } |
| 352 | 1 | equemene | Uptr = Mptr( Uptr, nn, 0, LDU );
|
| 353 | 1 | equemene | Aptr = Mptr( Aptr, 0, nn, lda ); nq0 += nn;
|
| 354 | 1 | equemene | |
| 355 | 1 | equemene | (void) HPL_bcast( PBCST, &test );
|
| 356 | 1 | equemene | } |
| 357 | 1 | equemene | /*
|
| 358 | 1 | equemene | * The panel has been forwarded at that point, finish the update
|
| 359 | 1 | equemene | */
|
| 360 | 1 | equemene | if( ( nn = n - nq0 ) > 0 ) |
| 361 | 1 | equemene | {
|
| 362 | 1 | equemene | HPL_dtrsm( HplColumnMajor, HplRight, HplLower, HplTrans, |
| 363 | 1 | equemene | HplUnit, nn, jb, HPL_rone, L1ptr, jb, Uptr, LDU ); |
| 364 | 1 | equemene | |
| 365 | 1 | equemene | if( curr != 0 ) |
| 366 | 1 | equemene | {
|
| 367 | 1 | equemene | #ifdef HPL_CALL_VSIPL
|
| 368 | 1 | equemene | /*
|
| 369 | 1 | equemene | * Create the matrix subviews
|
| 370 | 1 | equemene | */
|
| 371 | 1 | equemene | Uv1 = vsip_msubview_d( Uv0, nq0, 0, nn, jb );
|
| 372 | 1 | equemene | Av1 = vsip_msubview_d( Av0, PANEL->ii+jb, PANEL->jj+nq0, mp, nn ); |
| 373 | 1 | equemene | |
| 374 | 1 | equemene | vsip_gemp_d( -HPL_rone, Lv1, VSIP_MAT_NTRANS, Uv1, VSIP_MAT_TRANS, |
| 375 | 1 | equemene | HPL_rone, Av1 ); |
| 376 | 1 | equemene | /*
|
| 377 | 1 | equemene | * Destroy the matrix subviews
|
| 378 | 1 | equemene | */
|
| 379 | 1 | equemene | (void) vsip_mdestroy_d( Av1 );
|
| 380 | 1 | equemene | (void) vsip_mdestroy_d( Uv1 );
|
| 381 | 1 | equemene | #else
|
| 382 | 1 | equemene | HPL_dgemm( HplColumnMajor, HplNoTrans, HplTrans, mp, nn, |
| 383 | 1 | equemene | jb, -HPL_rone, L2ptr, ldl2, Uptr, LDU, HPL_rone, |
| 384 | 1 | equemene | Mptr( Aptr, jb, 0, lda ), lda );
|
| 385 | 1 | equemene | #endif
|
| 386 | 1 | equemene | HPL_dlatcpy( jb, nn, Uptr, LDU, Aptr, lda ); |
| 387 | 1 | equemene | } |
| 388 | 1 | equemene | else
|
| 389 | 1 | equemene | {
|
| 390 | 1 | equemene | #ifdef HPL_CALL_VSIPL
|
| 391 | 1 | equemene | /*
|
| 392 | 1 | equemene | * Create the matrix subviews
|
| 393 | 1 | equemene | */
|
| 394 | 1 | equemene | Uv1 = vsip_msubview_d( Uv0, nq0, 0, nn, jb );
|
| 395 | 1 | equemene | Av1 = vsip_msubview_d( Av0, PANEL->ii, PANEL->jj+nq0, mp, nn ); |
| 396 | 1 | equemene | |
| 397 | 1 | equemene | vsip_gemp_d( -HPL_rone, Lv1, VSIP_MAT_NTRANS, Uv1, VSIP_MAT_TRANS, |
| 398 | 1 | equemene | HPL_rone, Av1 ); |
| 399 | 1 | equemene | /*
|
| 400 | 1 | equemene | * Destroy the matrix subviews
|
| 401 | 1 | equemene | */
|
| 402 | 1 | equemene | (void) vsip_mdestroy_d( Av1 );
|
| 403 | 1 | equemene | (void) vsip_mdestroy_d( Uv1 );
|
| 404 | 1 | equemene | #else
|
| 405 | 1 | equemene | HPL_dgemm( HplColumnMajor, HplNoTrans, HplTrans, mp, nn, |
| 406 | 1 | equemene | jb, -HPL_rone, L2ptr, ldl2, Uptr, LDU, HPL_rone, |
| 407 | 1 | equemene | Aptr, lda ); |
| 408 | 1 | equemene | #endif
|
| 409 | 1 | equemene | } |
| 410 | 1 | equemene | } |
| 411 | 1 | equemene | #ifdef HPL_CALL_VSIPL
|
| 412 | 1 | equemene | /*
|
| 413 | 1 | equemene | * Destroy the matrix subviews
|
| 414 | 1 | equemene | */
|
| 415 | 1 | equemene | (void) vsip_mdestroy_d( Lv1 );
|
| 416 | 1 | equemene | /*
|
| 417 | 1 | equemene | * Release the blocks
|
| 418 | 1 | equemene | */
|
| 419 | 1 | equemene | (void) vsip_blockrelease_d( vsip_mgetblock_d( Uv0 ), VSIP_TRUE );
|
| 420 | 1 | equemene | (void) vsip_blockrelease_d( vsip_mgetblock_d( Lv0 ), VSIP_TRUE );
|
| 421 | 1 | equemene | (void) vsip_blockrelease_d( vsip_mgetblock_d( Av0 ), VSIP_TRUE );
|
| 422 | 1 | equemene | /*
|
| 423 | 1 | equemene | * Destroy the matrix views
|
| 424 | 1 | equemene | */
|
| 425 | 1 | equemene | (void) vsip_mdestroy_d( Uv0 );
|
| 426 | 1 | equemene | (void) vsip_mdestroy_d( Lv0 );
|
| 427 | 1 | equemene | (void) vsip_mdestroy_d( Av0 );
|
| 428 | 1 | equemene | #endif
|
| 429 | 1 | equemene | } |
| 430 | 1 | equemene | |
| 431 | 1 | equemene | PANEL->A = Mptr( PANEL->A, 0, n, lda ); PANEL->nq -= n; PANEL->jj += n;
|
| 432 | 1 | equemene | /*
|
| 433 | 1 | equemene | * return the outcome of the probe (should always be HPL_SUCCESS, the
|
| 434 | 1 | equemene | * panel broadcast is enforced in that routine).
|
| 435 | 1 | equemene | */
|
| 436 | 1 | equemene | if( PBCST != NULL ) *IFLAG = test; |
| 437 | 1 | equemene | #ifdef HPL_DETAILED_TIMING
|
| 438 | 1 | equemene | HPL_ptimer( HPL_TIMING_UPDATE ); |
| 439 | 1 | equemene | #endif
|
| 440 | 1 | equemene | /*
|
| 441 | 1 | equemene | * End of HPL_pdupdateNT
|
| 442 | 1 | equemene | */
|
| 443 | 1 | equemene | } |