root / src / pfact / HPL_pdpanllN.c @ 9
Historique | Voir | Annoter | Télécharger (9,88 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_pdpanllN
|
| 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_pdpanllN
|
| 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_pdpanllN factorizes a panel of columns that is a sub-array of a
|
| 76 | 1 | equemene | * larger one-dimensional panel A using the Left-looking variant of the
|
| 77 | 1 | equemene | * usual one-dimensional algorithm. The lower triangular N0-by-N0 upper
|
| 78 | 1 | equemene | * block of the panel is stored in no-transpose form (i.e. just like the
|
| 79 | 1 | equemene | * input matrix itself).
|
| 80 | 1 | equemene | *
|
| 81 | 1 | equemene | * Bi-directional exchange is used to perform the swap::broadcast
|
| 82 | 1 | equemene | * operations at once for one column in the panel. This results in a
|
| 83 | 1 | equemene | * lower number of slightly larger messages than usual. On P processes
|
| 84 | 1 | equemene | * and assuming bi-directional links, the running time of this function
|
| 85 | 1 | equemene | * can be approximated by (when N is equal to N0):
|
| 86 | 1 | equemene | *
|
| 87 | 1 | equemene | * N0 * log_2( P ) * ( lat + ( 2*N0 + 4 ) / bdwth ) +
|
| 88 | 1 | equemene | * N0^2 * ( M - N0/3 ) * gam2-3
|
| 89 | 1 | equemene | *
|
| 90 | 1 | equemene | * where M is the local number of rows of the panel, lat and bdwth are
|
| 91 | 1 | equemene | * the latency and bandwidth of the network for double precision real
|
| 92 | 1 | equemene | * words, and gam2-3 is an estimate of the Level 2 and Level 3 BLAS
|
| 93 | 1 | equemene | * rate of execution. The recursive algorithm allows indeed to almost
|
| 94 | 1 | equemene | * achieve Level 3 BLAS performance in the panel factorization. On a
|
| 95 | 1 | equemene | * large number of modern machines, this operation is however latency
|
| 96 | 1 | equemene | * bound, meaning that its cost can be estimated by only the latency
|
| 97 | 1 | equemene | * portion N0 * log_2(P) * lat. Mono-directional links will double this
|
| 98 | 1 | equemene | * communication cost.
|
| 99 | 1 | equemene | *
|
| 100 | 1 | equemene | * Note that one iteration of the the main loop is unrolled. The local
|
| 101 | 1 | equemene | * computation of the absolute value max of the next column is performed
|
| 102 | 1 | equemene | * just after its update by the current column. This allows to bring the
|
| 103 | 1 | equemene | * current column only once through cache at each step. The current
|
| 104 | 1 | equemene | * implementation does not perform any blocking for this sequence of
|
| 105 | 1 | equemene | * BLAS operations, however the design allows for plugging in an optimal
|
| 106 | 1 | equemene | * (machine-specific) specialized BLAS-like kernel. This idea has been
|
| 107 | 1 | equemene | * suggested to us by Fred Gustavson, IBM T.J. Watson Research Center.
|
| 108 | 1 | equemene | *
|
| 109 | 1 | equemene | * Arguments
|
| 110 | 1 | equemene | * =========
|
| 111 | 1 | equemene | *
|
| 112 | 1 | equemene | * PANEL (local input/output) HPL_T_panel *
|
| 113 | 1 | equemene | * On entry, PANEL points to the data structure containing the
|
| 114 | 1 | equemene | * panel information.
|
| 115 | 1 | equemene | *
|
| 116 | 1 | equemene | * M (local input) const int
|
| 117 | 1 | equemene | * On entry, M specifies the local number of rows of sub(A).
|
| 118 | 1 | equemene | *
|
| 119 | 1 | equemene | * N (local input) const int
|
| 120 | 1 | equemene | * On entry, N specifies the local number of columns of sub(A).
|
| 121 | 1 | equemene | *
|
| 122 | 1 | equemene | * ICOFF (global input) const int
|
| 123 | 1 | equemene | * On entry, ICOFF specifies the row and column offset of sub(A)
|
| 124 | 1 | equemene | * in A.
|
| 125 | 1 | equemene | *
|
| 126 | 1 | equemene | * WORK (local workspace) double *
|
| 127 | 1 | equemene | * On entry, WORK is a workarray of size at least 2*(4+2*N0).
|
| 128 | 1 | equemene | *
|
| 129 | 1 | equemene | * ---------------------------------------------------------------------
|
| 130 | 1 | equemene | */
|
| 131 | 1 | equemene | /*
|
| 132 | 1 | equemene | * .. Local Variables ..
|
| 133 | 1 | equemene | */
|
| 134 | 1 | equemene | double * A, * L1, * L1ptr;
|
| 135 | 1 | equemene | #ifdef HPL_CALL_VSIPL
|
| 136 | 1 | equemene | vsip_mview_d * Av0, * Av1, * Yv1, * Xv0, * Xv1; |
| 137 | 1 | equemene | #endif
|
| 138 | 1 | equemene | int Mm1, Nm1, curr, ii, iip1, jj, kk, lda,
|
| 139 | 1 | equemene | m=M, n0; |
| 140 | 1 | equemene | /* ..
|
| 141 | 1 | equemene | * .. Executable Statements ..
|
| 142 | 1 | equemene | */
|
| 143 | 1 | equemene | #ifdef HPL_DETAILED_TIMING
|
| 144 | 1 | equemene | HPL_ptimer( HPL_TIMING_PFACT ); |
| 145 | 1 | equemene | #endif
|
| 146 | 1 | equemene | A = PANEL->A; lda = PANEL->lda; |
| 147 | 1 | equemene | L1 = PANEL->L1; n0 = PANEL->jb; |
| 148 | 1 | equemene | curr = (int)( PANEL->grid->myrow == PANEL->prow );
|
| 149 | 1 | equemene | |
| 150 | 1 | equemene | Nm1 = N - 1; jj = ICOFF;
|
| 151 | 1 | equemene | if( curr != 0 ) { ii = ICOFF; iip1 = ii+1; Mm1 = m-1; } |
| 152 | 1 | equemene | else { ii = 0; iip1 = ii; Mm1 = m; } |
| 153 | 1 | equemene | #ifdef HPL_CALL_VSIPL
|
| 154 | 1 | equemene | /*
|
| 155 | 1 | equemene | * Admit the blocks
|
| 156 | 1 | equemene | */
|
| 157 | 1 | equemene | (void) vsip_blockadmit_d( PANEL->Ablock, VSIP_TRUE );
|
| 158 | 1 | equemene | (void) vsip_blockadmit_d( PANEL->L1block, VSIP_TRUE );
|
| 159 | 1 | equemene | /*
|
| 160 | 1 | equemene | * Create the matrix views
|
| 161 | 1 | equemene | */
|
| 162 | 1 | equemene | Av0 = vsip_mbind_d( PANEL->Ablock, 0, 1, lda, lda, PANEL->pmat->nq ); |
| 163 | 1 | equemene | Xv0 = vsip_mbind_d( PANEL->L1block, 0, 1, PANEL->jb, PANEL->jb, PANEL->jb ); |
| 164 | 1 | equemene | #endif
|
| 165 | 1 | equemene | /*
|
| 166 | 1 | equemene | * Find local absolute value max in first column and initialize WORK[0:3]
|
| 167 | 1 | equemene | */
|
| 168 | 1 | equemene | HPL_dlocmax( PANEL, m, ii, jj, WORK ); |
| 169 | 1 | equemene | |
| 170 | 1 | equemene | while( Nm1 > 0 ) |
| 171 | 1 | equemene | {
|
| 172 | 1 | equemene | /*
|
| 173 | 1 | equemene | * Swap and broadcast the current row
|
| 174 | 1 | equemene | */
|
| 175 | 1 | equemene | HPL_pdmxswp( PANEL, m, ii, jj, WORK ); |
| 176 | 1 | equemene | HPL_dlocswpN( PANEL, ii, jj, WORK ); |
| 177 | 1 | equemene | |
| 178 | 1 | equemene | L1ptr = Mptr( L1, ICOFF, jj+1, n0 ); kk = jj + 1 - ICOFF; |
| 179 | 1 | equemene | HPL_dtrsv( HplColumnMajor, HplLower, HplNoTrans, HplUnit, kk, |
| 180 | 1 | equemene | Mptr( L1, ICOFF, ICOFF, n0 ), n0, L1ptr, 1 );
|
| 181 | 1 | equemene | /*
|
| 182 | 1 | equemene | * Scale current column by its absolute value max entry - Update and
|
| 183 | 1 | equemene | * find local absolute value max in next column (Only one pass through
|
| 184 | 1 | equemene | * cache for each next column). This sequence of operations could bene-
|
| 185 | 1 | equemene | * fit from a specialized blocked implementation.
|
| 186 | 1 | equemene | */
|
| 187 | 1 | equemene | if( WORK[0] != HPL_rzero ) |
| 188 | 1 | equemene | HPL_dscal( Mm1, HPL_rone / WORK[0], Mptr( A, iip1, jj, lda ), 1 ); |
| 189 | 1 | equemene | #ifdef HPL_CALL_VSIPL
|
| 190 | 1 | equemene | /*
|
| 191 | 1 | equemene | * Create the matrix subviews
|
| 192 | 1 | equemene | */
|
| 193 | 1 | equemene | Av1 = vsip_msubview_d( Av0, PANEL->ii+iip1, PANEL->jj+ICOFF, Mm1, kk ); |
| 194 | 1 | equemene | Xv1 = vsip_msubview_d( Xv0, ICOFF, jj+1, kk, 1 ); |
| 195 | 1 | equemene | Yv1 = vsip_msubview_d( Av0, PANEL->ii+iip1, PANEL->jj+jj+1, Mm1, 1 ); |
| 196 | 1 | equemene | |
| 197 | 1 | equemene | vsip_gemp_d( -HPL_rone, Av1, VSIP_MAT_NTRANS, Xv1, VSIP_MAT_NTRANS, |
| 198 | 1 | equemene | HPL_rone, Yv1 ); |
| 199 | 1 | equemene | /*
|
| 200 | 1 | equemene | * Destroy the matrix subviews
|
| 201 | 1 | equemene | */
|
| 202 | 1 | equemene | (void) vsip_mdestroy_d( Yv1 );
|
| 203 | 1 | equemene | (void) vsip_mdestroy_d( Xv1 );
|
| 204 | 1 | equemene | (void) vsip_mdestroy_d( Av1 );
|
| 205 | 1 | equemene | #else
|
| 206 | 1 | equemene | HPL_dgemv( HplColumnMajor, HplNoTrans, Mm1, kk, -HPL_rone, |
| 207 | 1 | equemene | Mptr( A, iip1, ICOFF, lda ), lda, L1ptr, 1,
|
| 208 | 1 | equemene | HPL_rone, Mptr( A, iip1, jj+1, lda ), 1 ); |
| 209 | 1 | equemene | #endif
|
| 210 | 1 | equemene | HPL_dlocmax( PANEL, Mm1, iip1, jj+1, WORK );
|
| 211 | 1 | equemene | if( curr != 0 ) |
| 212 | 1 | equemene | {
|
| 213 | 1 | equemene | HPL_dcopy( kk, L1ptr, 1, Mptr( A, ICOFF, jj+1, lda ), 1 ); |
| 214 | 1 | equemene | ii = iip1; iip1++; m = Mm1; Mm1--; |
| 215 | 1 | equemene | } |
| 216 | 1 | equemene | Nm1--; jj++; |
| 217 | 1 | equemene | } |
| 218 | 1 | equemene | /*
|
| 219 | 1 | equemene | * Swap and broadcast last row - Scale last column by its absolute value
|
| 220 | 1 | equemene | * max entry
|
| 221 | 1 | equemene | */
|
| 222 | 1 | equemene | HPL_pdmxswp( PANEL, m, ii, jj, WORK ); |
| 223 | 1 | equemene | HPL_dlocswpN( PANEL, ii, jj, WORK ); |
| 224 | 1 | equemene | if( WORK[0] != HPL_rzero ) |
| 225 | 1 | equemene | HPL_dscal( Mm1, HPL_rone / WORK[0], Mptr( A, iip1, jj, lda ), 1 ); |
| 226 | 1 | equemene | #ifdef HPL_CALL_VSIPL
|
| 227 | 1 | equemene | /*
|
| 228 | 1 | equemene | * Release the blocks
|
| 229 | 1 | equemene | */
|
| 230 | 1 | equemene | (void) vsip_blockrelease_d( vsip_mgetblock_d( Xv0 ), VSIP_TRUE );
|
| 231 | 1 | equemene | (void) vsip_blockrelease_d( vsip_mgetblock_d( Av0 ), VSIP_TRUE );
|
| 232 | 1 | equemene | /*
|
| 233 | 1 | equemene | * Destroy the matrix views
|
| 234 | 1 | equemene | */
|
| 235 | 1 | equemene | (void) vsip_mdestroy_d( Xv0 );
|
| 236 | 1 | equemene | (void) vsip_mdestroy_d( Av0 );
|
| 237 | 1 | equemene | #endif
|
| 238 | 1 | equemene | #ifdef HPL_DETAILED_TIMING
|
| 239 | 1 | equemene | HPL_ptimer( HPL_TIMING_PFACT ); |
| 240 | 1 | equemene | #endif
|
| 241 | 1 | equemene | /*
|
| 242 | 1 | equemene | * End of HPL_pdpanllN
|
| 243 | 1 | equemene | */
|
| 244 | 1 | equemene | } |