root / src / pfact / HPL_pdrpanllT.c @ 1
Historique | Voir | Annoter | Télécharger (9,82 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_pdrpanllT
|
| 54 | 1 | equemene | ( |
| 55 | 1 | equemene | HPL_T_panel * PANEL, |
| 56 | 1 | equemene | const int M, |
| 57 | 1 | equemene | const int N, |
| 58 | 1 | equemene | const int ICOFF, |
| 59 | 1 | equemene | double * WORK
|
| 60 | 1 | equemene | ) |
| 61 | 1 | equemene | #else
|
| 62 | 1 | equemene | void HPL_pdrpanllT
|
| 63 | 1 | equemene | ( PANEL, M, N, ICOFF, WORK ) |
| 64 | 1 | equemene | HPL_T_panel * PANEL; |
| 65 | 1 | equemene | const int M; |
| 66 | 1 | equemene | const int N; |
| 67 | 1 | equemene | const int ICOFF; |
| 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_pdrpanllT recursively factorizes a panel of columns using the
|
| 76 | 1 | equemene | * recursive Left-looking variant of the one-dimensional algorithm. The
|
| 77 | 1 | equemene | * lower triangular N0-by-N0 upper block of the panel is stored in
|
| 78 | 1 | equemene | * transpose form.
|
| 79 | 1 | equemene | *
|
| 80 | 1 | equemene | * Bi-directional exchange is used to perform the swap::broadcast
|
| 81 | 1 | equemene | * operations at once for one column in the panel. This results in a
|
| 82 | 1 | equemene | * lower number of slightly larger messages than usual. On P processes
|
| 83 | 1 | equemene | * and assuming bi-directional links, the running time of this function
|
| 84 | 1 | equemene | * can be approximated by (when N is equal to N0):
|
| 85 | 1 | equemene | *
|
| 86 | 1 | equemene | * N0 * log_2( P ) * ( lat + ( 2*N0 + 4 ) / bdwth ) +
|
| 87 | 1 | equemene | * N0^2 * ( M - N0/3 ) * gam2-3
|
| 88 | 1 | equemene | *
|
| 89 | 1 | equemene | * where M is the local number of rows of the panel, lat and bdwth are
|
| 90 | 1 | equemene | * the latency and bandwidth of the network for double precision real
|
| 91 | 1 | equemene | * words, and gam2-3 is an estimate of the Level 2 and Level 3 BLAS
|
| 92 | 1 | equemene | * rate of execution. The recursive algorithm allows indeed to almost
|
| 93 | 1 | equemene | * achieve Level 3 BLAS performance in the panel factorization. On a
|
| 94 | 1 | equemene | * large number of modern machines, this operation is however latency
|
| 95 | 1 | equemene | * bound, meaning that its cost can be estimated by only the latency
|
| 96 | 1 | equemene | * portion N0 * log_2(P) * lat. Mono-directional links will double this
|
| 97 | 1 | equemene | * communication cost.
|
| 98 | 1 | equemene | *
|
| 99 | 1 | equemene | * Arguments
|
| 100 | 1 | equemene | * =========
|
| 101 | 1 | equemene | *
|
| 102 | 1 | equemene | * PANEL (local input/output) HPL_T_panel *
|
| 103 | 1 | equemene | * On entry, PANEL points to the data structure containing the
|
| 104 | 1 | equemene | * panel information.
|
| 105 | 1 | equemene | *
|
| 106 | 1 | equemene | * M (local input) const int
|
| 107 | 1 | equemene | * On entry, M specifies the local number of rows of sub(A).
|
| 108 | 1 | equemene | *
|
| 109 | 1 | equemene | * N (local input) const int
|
| 110 | 1 | equemene | * On entry, N specifies the local number of columns of sub(A).
|
| 111 | 1 | equemene | *
|
| 112 | 1 | equemene | * ICOFF (global input) const int
|
| 113 | 1 | equemene | * On entry, ICOFF specifies the row and column offset of sub(A)
|
| 114 | 1 | equemene | * in A.
|
| 115 | 1 | equemene | *
|
| 116 | 1 | equemene | * WORK (local workspace) double *
|
| 117 | 1 | equemene | * On entry, WORK is a workarray of size at least 2*(4+2*N0).
|
| 118 | 1 | equemene | *
|
| 119 | 1 | equemene | * ---------------------------------------------------------------------
|
| 120 | 1 | equemene | */
|
| 121 | 1 | equemene | /*
|
| 122 | 1 | equemene | * .. Local Variables ..
|
| 123 | 1 | equemene | */
|
| 124 | 1 | equemene | double * A, * Aptr, * L1, * L1ptr;
|
| 125 | 1 | equemene | #ifdef HPL_CALL_VSIPL
|
| 126 | 1 | equemene | vsip_mview_d * Av0, * Lv0, * Av1, * Av2, * Lv1; |
| 127 | 1 | equemene | #endif
|
| 128 | 1 | equemene | int curr, ii, ioff, jb, jj, lda, m, n, n0, nb,
|
| 129 | 1 | equemene | nbdiv, nbmin; |
| 130 | 1 | equemene | /* ..
|
| 131 | 1 | equemene | * .. Executable Statements ..
|
| 132 | 1 | equemene | */
|
| 133 | 1 | equemene | if( N <= ( nbmin = PANEL->algo->nbmin ) )
|
| 134 | 1 | equemene | { PANEL->algo->pffun( PANEL, M, N, ICOFF, WORK ); return; }
|
| 135 | 1 | equemene | /*
|
| 136 | 1 | equemene | * Find new recursive blocking factor. To avoid an infinite loop, one
|
| 137 | 1 | equemene | * must guarantee: 1 <= jb < N, knowing that N is greater than NBMIN.
|
| 138 | 1 | equemene | * First, we compute nblocks: the number of blocks of size NBMIN in N,
|
| 139 | 1 | equemene | * including the last one that may be smaller. nblocks is thus larger
|
| 140 | 1 | equemene | * than or equal to one, since N >= NBMIN.
|
| 141 | 1 | equemene | * The ratio ( nblocks + NDIV - 1 ) / NDIV is thus larger than or equal
|
| 142 | 1 | equemene | * to one as well. For NDIV >= 2, we are guaranteed that the quan-
|
| 143 | 1 | equemene | * tity ( ( nblocks + NDIV - 1 ) / NDIV ) * NBMIN is less than N and
|
| 144 | 1 | equemene | * greater than or equal to NBMIN.
|
| 145 | 1 | equemene | */
|
| 146 | 1 | equemene | nbdiv = PANEL->algo->nbdiv; ii = jj = 0; m = M; n = N;
|
| 147 | 1 | equemene | nb = jb = ( (((N+nbmin-1) / nbmin) + nbdiv - 1) / nbdiv ) * nbmin; |
| 148 | 1 | equemene | |
| 149 | 1 | equemene | A = PANEL->A; lda = PANEL->lda; |
| 150 | 1 | equemene | L1 = PANEL->L1; n0 = PANEL->jb; |
| 151 | 1 | equemene | L1ptr = Mptr( L1, ICOFF, ICOFF, n0 ); |
| 152 | 1 | equemene | curr = (int)( PANEL->grid->myrow == PANEL->prow );
|
| 153 | 1 | equemene | |
| 154 | 1 | equemene | if( curr != 0 ) Aptr = Mptr( A, ICOFF, ICOFF, lda ); |
| 155 | 1 | equemene | else Aptr = Mptr( A, 0, ICOFF, lda ); |
| 156 | 1 | equemene | /*
|
| 157 | 1 | equemene | * The triangular solve is replicated in every process row. The panel
|
| 158 | 1 | equemene | * factorization is such that the first rows of A are accumulated in
|
| 159 | 1 | equemene | * every process row during the (panel) swapping phase. We ensure this
|
| 160 | 1 | equemene | * way a minimum amount of communication during the entire panel facto-
|
| 161 | 1 | equemene | * rization.
|
| 162 | 1 | equemene | */
|
| 163 | 1 | equemene | do
|
| 164 | 1 | equemene | {
|
| 165 | 1 | equemene | n -= jb; ioff = ICOFF + jj; |
| 166 | 1 | equemene | /*
|
| 167 | 1 | equemene | * Replicated solve - Local update - Factor current panel
|
| 168 | 1 | equemene | */
|
| 169 | 1 | equemene | HPL_dtrsm( HplColumnMajor, HplRight, HplUpper, HplNoTrans, |
| 170 | 1 | equemene | HplUnit, jb, jj, HPL_rone, L1ptr, n0, Mptr( L1ptr, |
| 171 | 1 | equemene | jj, 0, n0 ), n0 );
|
| 172 | 1 | equemene | #ifdef HPL_CALL_VSIPL
|
| 173 | 1 | equemene | /*
|
| 174 | 1 | equemene | * Admit the blocks
|
| 175 | 1 | equemene | */
|
| 176 | 1 | equemene | (void) vsip_blockadmit_d( PANEL->Ablock, VSIP_TRUE );
|
| 177 | 1 | equemene | (void) vsip_blockadmit_d( PANEL->L1block, VSIP_TRUE );
|
| 178 | 1 | equemene | /*
|
| 179 | 1 | equemene | * Create the matrix views
|
| 180 | 1 | equemene | */
|
| 181 | 1 | equemene | Av0 = vsip_mbind_d( PANEL->Ablock, 0, 1, lda, lda, PANEL->pmat->nq ); |
| 182 | 1 | equemene | Lv0 = vsip_mbind_d( PANEL->L1block, 0, 1, n0, n0, n0 ); |
| 183 | 1 | equemene | /*
|
| 184 | 1 | equemene | * Create the matrix subviews
|
| 185 | 1 | equemene | */
|
| 186 | 1 | equemene | if( curr != 0 ) |
| 187 | 1 | equemene | {
|
| 188 | 1 | equemene | Av1 = vsip_msubview_d( Av0, PANEL->ii+ICOFF+ii, PANEL->jj+ICOFF, |
| 189 | 1 | equemene | m, jj ); |
| 190 | 1 | equemene | Av2 = vsip_msubview_d( Av0, PANEL->ii+ICOFF+ii, PANEL->jj+ioff, |
| 191 | 1 | equemene | m, jj ); |
| 192 | 1 | equemene | } |
| 193 | 1 | equemene | else
|
| 194 | 1 | equemene | {
|
| 195 | 1 | equemene | Av1 = vsip_msubview_d( Av0, PANEL->ii+ii, PANEL->jj+ICOFF, m, jj ); |
| 196 | 1 | equemene | Av2 = vsip_msubview_d( Av0, PANEL->ii+ii, PANEL->jj+ioff, m, jj ); |
| 197 | 1 | equemene | } |
| 198 | 1 | equemene | Lv1 = vsip_msubview_d( Lv0, ioff, ICOFF, jb, jj ); |
| 199 | 1 | equemene | |
| 200 | 1 | equemene | vsip_gemp_d( -HPL_rone, Av1, VSIP_MAT_NTRANS, Lv1, VSIP_MAT_TRANS, |
| 201 | 1 | equemene | HPL_rone, Av2 ); |
| 202 | 1 | equemene | /*
|
| 203 | 1 | equemene | * Destroy the matrix subviews
|
| 204 | 1 | equemene | */
|
| 205 | 1 | equemene | (void) vsip_mdestroy_d( Lv1 );
|
| 206 | 1 | equemene | (void) vsip_mdestroy_d( Av1 );
|
| 207 | 1 | equemene | (void) vsip_mdestroy_d( Av2 );
|
| 208 | 1 | equemene | /*
|
| 209 | 1 | equemene | * Release the blocks
|
| 210 | 1 | equemene | */
|
| 211 | 1 | equemene | (void) vsip_blockrelease_d( vsip_mgetblock_d( Lv0 ), VSIP_TRUE );
|
| 212 | 1 | equemene | (void) vsip_blockrelease_d( vsip_mgetblock_d( Av0 ), VSIP_TRUE );
|
| 213 | 1 | equemene | /*
|
| 214 | 1 | equemene | * Destroy the matrix views
|
| 215 | 1 | equemene | */
|
| 216 | 1 | equemene | (void) vsip_mdestroy_d( Lv0 );
|
| 217 | 1 | equemene | (void) vsip_mdestroy_d( Av0 );
|
| 218 | 1 | equemene | #else
|
| 219 | 1 | equemene | HPL_dgemm( HplColumnMajor, HplNoTrans, HplTrans, m, jb, |
| 220 | 1 | equemene | jj, -HPL_rone, Mptr( Aptr, ii, 0, lda ), lda,
|
| 221 | 1 | equemene | Mptr( L1ptr, jj, 0, n0 ), n0, HPL_rone,
|
| 222 | 1 | equemene | Mptr( Aptr, ii, jj, lda ), lda ); |
| 223 | 1 | equemene | #endif
|
| 224 | 1 | equemene | HPL_pdrpanllT( PANEL, m, jb, ioff, WORK ); |
| 225 | 1 | equemene | /*
|
| 226 | 1 | equemene | * Copy back upper part of A in current process row - Go the next block
|
| 227 | 1 | equemene | */
|
| 228 | 1 | equemene | if( curr != 0 ) |
| 229 | 1 | equemene | {
|
| 230 | 1 | equemene | HPL_dlatcpy( ioff, jb, Mptr( L1, ioff, 0, n0 ), n0,
|
| 231 | 1 | equemene | Mptr( A, 0, ioff, lda ), lda );
|
| 232 | 1 | equemene | ii += jb; m -= jb; |
| 233 | 1 | equemene | } |
| 234 | 1 | equemene | jj += jb; jb = Mmin( n, nb ); |
| 235 | 1 | equemene | |
| 236 | 1 | equemene | } while( n > 0 ); |
| 237 | 1 | equemene | /*
|
| 238 | 1 | equemene | * End of HPL_pdrpanllT
|
| 239 | 1 | equemene | */
|
| 240 | 1 | equemene | } |