| 1 |
/* $Id$ */ |
| 2 |
|
| 3 |
/**************************************************************/ |
| 4 |
|
| 5 |
/* Paso: RILU preconditioner with reordering */ |
| 6 |
|
| 7 |
/**************************************************************/ |
| 8 |
|
| 9 |
/* Copyrights by ACcESS Australia 2003,2004,2005 */ |
| 10 |
/* Author: gross@access.edu.au */ |
| 11 |
|
| 12 |
/**************************************************************/ |
| 13 |
|
| 14 |
#include "Paso.h" |
| 15 |
#include "Solver.h" |
| 16 |
#include "PasoUtil.h" |
| 17 |
|
| 18 |
/**************************************************************/ |
| 19 |
|
| 20 |
/* free all memory used by RILU */ |
| 21 |
|
| 22 |
void Paso_Solver_RILU_free(Paso_Solver_RILU * in) { |
| 23 |
if (in!=NULL) { |
| 24 |
Paso_Solver_RILU_free(in->RILU_of_Schur); |
| 25 |
MEMFREE(in->inv_A_FF); |
| 26 |
MEMFREE(in->A_FF_pivot); |
| 27 |
Paso_SystemMatrix_dealloc(in->A_FC); |
| 28 |
Paso_SystemMatrix_dealloc(in->A_CF); |
| 29 |
MEMFREE(in->rows_in_F); |
| 30 |
MEMFREE(in->rows_in_C); |
| 31 |
MEMFREE(in->mask_F); |
| 32 |
MEMFREE(in->mask_C); |
| 33 |
MEMFREE(in->x_F); |
| 34 |
MEMFREE(in->b_F); |
| 35 |
MEMFREE(in->x_C); |
| 36 |
MEMFREE(in->b_C); |
| 37 |
MEMFREE(in); |
| 38 |
} |
| 39 |
} |
| 40 |
|
| 41 |
/**************************************************************/ |
| 42 |
|
| 43 |
/* constructs the block-block factorization of |
| 44 |
|
| 45 |
[ A_FF A_FC ] |
| 46 |
A_p= |
| 47 |
[ A_CF A_FF ] |
| 48 |
|
| 49 |
to |
| 50 |
|
| 51 |
[ I 0 ] [ A_FF 0 ] [ I invA_FF*A_FF ] |
| 52 |
[ A_CF*invA_FF I ] [ 0 S ] [ 0 I ] |
| 53 |
|
| 54 |
|
| 55 |
where S=A_FF-ACF*invA_FF*A_FC within the shape of S |
| 56 |
|
| 57 |
then RILU is applied to S again until S becomes empty |
| 58 |
|
| 59 |
*/ |
| 60 |
Paso_Solver_RILU* Paso_Solver_getRILU(Paso_SystemMatrix * A_p,bool_t verbose) { |
| 61 |
Paso_Solver_RILU* out=NULL; |
| 62 |
dim_t n=A_p->num_rows; |
| 63 |
dim_t n_block=A_p->row_block_size; |
| 64 |
index_t* mis_marker=NULL; |
| 65 |
index_t* counter=NULL; |
| 66 |
index_t iPtr,*index, *where_p; |
| 67 |
dim_t i,k; |
| 68 |
Paso_SystemMatrix * schur=NULL; |
| 69 |
double A11,A12,A13,A21,A22,A23,A31,A32,A33,D,time0,time1,time2; |
| 70 |
|
| 71 |
|
| 72 |
/* identify independend set of rows/columns */ |
| 73 |
mis_marker=TMPMEMALLOC(n,index_t); |
| 74 |
counter=TMPMEMALLOC(n,index_t); |
| 75 |
out=MEMALLOC(1,Paso_Solver_RILU); |
| 76 |
out->RILU_of_Schur=NULL; |
| 77 |
out->inv_A_FF=NULL; |
| 78 |
out->A_FF_pivot=NULL; |
| 79 |
out->A_FC=NULL; |
| 80 |
out->A_CF=NULL; |
| 81 |
out->rows_in_F=NULL; |
| 82 |
out->rows_in_C=NULL; |
| 83 |
out->mask_F=NULL; |
| 84 |
out->mask_C=NULL; |
| 85 |
out->x_F=NULL; |
| 86 |
out->b_F=NULL; |
| 87 |
out->x_C=NULL; |
| 88 |
out->b_C=NULL; |
| 89 |
|
| 90 |
if ( !(Paso_checkPtr(mis_marker) || Paso_checkPtr(out) || Paso_checkPtr(counter) ) ) { |
| 91 |
/* identify independend set of rows/columns */ |
| 92 |
time0=Paso_timer(); |
| 93 |
#pragma omp parallel for private(i) schedule(static) |
| 94 |
for (i=0;i<n;++i) mis_marker[i]=-1; |
| 95 |
Paso_SystemMatrixPattern_mis(A_p->pattern,mis_marker); |
| 96 |
time2=Paso_timer()-time0; |
| 97 |
if (Paso_noError()) { |
| 98 |
#pragma omp parallel for private(i) schedule(static) |
| 99 |
for (i = 0; i < n; ++i) counter[i]=mis_marker[i]; |
| 100 |
out->n=n; |
| 101 |
out->n_block=n_block; |
| 102 |
out->n_F=Paso_Util_cumsum(n,counter); |
| 103 |
out->mask_F=MEMALLOC(n,index_t); |
| 104 |
out->rows_in_F=MEMALLOC(out->n_F,index_t); |
| 105 |
out->inv_A_FF=MEMALLOC(n_block*n_block*out->n_F,double); |
| 106 |
out->A_FF_pivot=NULL; /* later use for block size>3 */ |
| 107 |
if (! (Paso_checkPtr(out->mask_F) || Paso_checkPtr(out->inv_A_FF) || Paso_checkPtr(out->rows_in_F) ) ) { |
| 108 |
#pragma omp parallel |
| 109 |
{ |
| 110 |
/* creates an index for F from mask */ |
| 111 |
#pragma omp for private(i) schedule(static) |
| 112 |
for (i = 0; i < out->n_F; ++i) out->rows_in_F[i]=-1; |
| 113 |
#pragma omp for private(i) schedule(static) |
| 114 |
for (i = 0; i < n; ++i) { |
| 115 |
if (mis_marker[i]) { |
| 116 |
out->rows_in_F[counter[i]]=i; |
| 117 |
out->mask_F[i]=counter[i]; |
| 118 |
} else { |
| 119 |
out->mask_F[i]=-1; |
| 120 |
} |
| 121 |
} |
| 122 |
#pragma omp for private(i, where_p,iPtr,A11,A12,A13,A21,A22,A23,A31,A32,A33,D,index) schedule(static) |
| 123 |
for (i = 0; i < out->n_F; i++) { |
| 124 |
/* find main diagonal */ |
| 125 |
iPtr=A_p->pattern->ptr[out->rows_in_F[i]]; |
| 126 |
index=&(A_p->pattern->index[iPtr]); |
| 127 |
where_p=(index_t*)bsearch(&out->rows_in_F[i], |
| 128 |
index, |
| 129 |
A_p->pattern->ptr[out->rows_in_F[i] + 1]-A_p->pattern->ptr[out->rows_in_F[i]], |
| 130 |
sizeof(index_t), |
| 131 |
Paso_comparIndex); |
| 132 |
if (where_p==NULL) { |
| 133 |
Paso_setError(VALUE_ERROR, "Paso_Solver_getRILU: main diagonal element missing."); |
| 134 |
} else { |
| 135 |
iPtr+=(index_t)(where_p-index); |
| 136 |
/* get inverse of A_FF block: */ |
| 137 |
if (n_block==1) { |
| 138 |
if (ABS(A_p->val[iPtr])>0.) { |
| 139 |
out->inv_A_FF[i]=1./A_p->val[iPtr]; |
| 140 |
} else { |
| 141 |
Paso_setError(ZERO_DIVISION_ERROR, "Paso_Solver_getRILU: Break-down in RILU decomposition: non-regular main diagonal block."); |
| 142 |
} |
| 143 |
} else if (n_block==2) { |
| 144 |
A11=A_p->val[iPtr*4]; |
| 145 |
A21=A_p->val[iPtr*4+1]; |
| 146 |
A12=A_p->val[iPtr*4+2]; |
| 147 |
A22=A_p->val[iPtr*4+3]; |
| 148 |
D = A11*A22-A12*A21; |
| 149 |
if (ABS(D) > 0 ){ |
| 150 |
D=1./D; |
| 151 |
out->inv_A_FF[i*4]= A22*D; |
| 152 |
out->inv_A_FF[i*4+1]=-A21*D; |
| 153 |
out->inv_A_FF[i*4+2]=-A12*D; |
| 154 |
out->inv_A_FF[i*4+3]= A11*D; |
| 155 |
} else { |
| 156 |
Paso_setError(ZERO_DIVISION_ERROR, "Paso_Solver_getRILU:Break-down in RILU decomposition: non-regular main diagonal block."); |
| 157 |
} |
| 158 |
} else if (n_block==3) { |
| 159 |
A11=A_p->val[iPtr*9 ]; |
| 160 |
A21=A_p->val[iPtr*9+1]; |
| 161 |
A31=A_p->val[iPtr*9+2]; |
| 162 |
A12=A_p->val[iPtr*9+3]; |
| 163 |
A22=A_p->val[iPtr*9+4]; |
| 164 |
A32=A_p->val[iPtr*9+5]; |
| 165 |
A13=A_p->val[iPtr*9+6]; |
| 166 |
A23=A_p->val[iPtr*9+7]; |
| 167 |
A33=A_p->val[iPtr*9+8]; |
| 168 |
D = A11*(A22*A33-A23*A32)+ A12*(A31*A23-A21*A33)+A13*(A21*A32-A31*A22); |
| 169 |
if (ABS(D) > 0 ){ |
| 170 |
D=1./D; |
| 171 |
out->inv_A_FF[i*9 ]=(A22*A33-A23*A32)*D; |
| 172 |
out->inv_A_FF[i*9+1]=(A31*A23-A21*A33)*D; |
| 173 |
out->inv_A_FF[i*9+2]=(A21*A32-A31*A22)*D; |
| 174 |
out->inv_A_FF[i*9+3]=(A13*A32-A12*A33)*D; |
| 175 |
out->inv_A_FF[i*9+4]=(A11*A33-A31*A13)*D; |
| 176 |
out->inv_A_FF[i*9+5]=(A12*A31-A11*A32)*D; |
| 177 |
out->inv_A_FF[i*9+6]=(A12*A23-A13*A22)*D; |
| 178 |
out->inv_A_FF[i*9+7]=(A13*A21-A11*A23)*D; |
| 179 |
out->inv_A_FF[i*9+8]=(A11*A22-A12*A21)*D; |
| 180 |
} else { |
| 181 |
Paso_setError(ZERO_DIVISION_ERROR, "Paso_Solver_getRILU:Break-down in RILU decomposition: non-regular main diagonal block."); |
| 182 |
} |
| 183 |
} |
| 184 |
} |
| 185 |
} |
| 186 |
} /* end parallel region */ |
| 187 |
|
| 188 |
if( Paso_noError()) { |
| 189 |
/* if there are no nodes in the coarse level there is no more work to do */ |
| 190 |
out->n_C=n-out->n_F; |
| 191 |
if (out->n_C>0) { |
| 192 |
out->rows_in_C=MEMALLOC(out->n_C,index_t); |
| 193 |
out->mask_C=MEMALLOC(n,index_t); |
| 194 |
if (! (Paso_checkPtr(out->mask_C) || Paso_checkPtr(out->rows_in_C) ) ) { |
| 195 |
/* creates an index for C from mask */ |
| 196 |
#pragma omp parallel for private(i) schedule(static) |
| 197 |
for (i = 0; i < n; ++i) counter[i]=! mis_marker[i]; |
| 198 |
Paso_Util_cumsum(n,counter); |
| 199 |
#pragma omp parallel |
| 200 |
{ |
| 201 |
#pragma omp for private(i) schedule(static) |
| 202 |
for (i = 0; i < out->n_C; ++i) out->rows_in_C[i]=-1; |
| 203 |
#pragma omp for private(i) schedule(static) |
| 204 |
for (i = 0; i < n; ++i) { |
| 205 |
if (! mis_marker[i]) { |
| 206 |
out->rows_in_C[counter[i]]=i; |
| 207 |
out->mask_C[i]=counter[i]; |
| 208 |
} else { |
| 209 |
out->mask_C[i]=-1; |
| 210 |
} |
| 211 |
} |
| 212 |
} /* end parallel region */ |
| 213 |
/* get A_CF block: */ |
| 214 |
out->A_CF=Paso_SystemMatrix_getSubmatrix(A_p,out->n_C,out->rows_in_C,out->mask_F); |
| 215 |
if (Paso_noError()) { |
| 216 |
/* get A_FC block: */ |
| 217 |
out->A_FC=Paso_SystemMatrix_getSubmatrix(A_p,out->n_F,out->rows_in_F,out->mask_C); |
| 218 |
/* get A_FF block: */ |
| 219 |
if (Paso_noError()) { |
| 220 |
schur=Paso_SystemMatrix_getSubmatrix(A_p,out->n_C,out->rows_in_C,out->mask_C); |
| 221 |
time0=Paso_timer()-time0; |
| 222 |
if (Paso_noError()) { |
| 223 |
time1=Paso_timer(); |
| 224 |
/* update A_CC block to get Schur complement and then apply RILU to it */ |
| 225 |
Paso_Solver_updateIncompleteSchurComplement(schur,out->A_CF,out->inv_A_FF,out->A_FF_pivot,out->A_FC); |
| 226 |
time1=Paso_timer()-time1; |
| 227 |
out->RILU_of_Schur=Paso_Solver_getRILU(schur,verbose); |
| 228 |
Paso_SystemMatrix_dealloc(schur); |
| 229 |
} |
| 230 |
/* allocate work arrays for RILU application */ |
| 231 |
if (Paso_noError()) { |
| 232 |
out->x_F=MEMALLOC(n_block*out->n_F,double); |
| 233 |
out->b_F=MEMALLOC(n_block*out->n_F,double); |
| 234 |
out->x_C=MEMALLOC(n_block*out->n_C,double); |
| 235 |
out->b_C=MEMALLOC(n_block*out->n_C,double); |
| 236 |
if (! (Paso_checkPtr(out->x_F) || Paso_checkPtr(out->b_F) || Paso_checkPtr(out->x_C) || Paso_checkPtr(out->b_C) ) ) { |
| 237 |
#pragma omp parallel |
| 238 |
{ |
| 239 |
#pragma omp for private(i,k) schedule(static) |
| 240 |
for (i = 0; i < out->n_F; ++i) { |
| 241 |
for (k=0; k<n_block;++k) { |
| 242 |
out->x_F[i*n_block+k]=0.; |
| 243 |
out->b_F[i*n_block+k]=0.; |
| 244 |
} |
| 245 |
} |
| 246 |
#pragma omp for private(i,k) schedule(static) |
| 247 |
for (i = 0; i < out->n_C; ++i) { |
| 248 |
for (k=0; k<n_block;++k) { |
| 249 |
out->x_C[i*n_block+k]=0.; |
| 250 |
out->b_C[i*n_block+k]=0.; |
| 251 |
} |
| 252 |
} |
| 253 |
} /* end parallel region */ |
| 254 |
} |
| 255 |
} |
| 256 |
} |
| 257 |
} |
| 258 |
} |
| 259 |
} |
| 260 |
} |
| 261 |
} |
| 262 |
} |
| 263 |
} |
| 264 |
TMPMEMFREE(mis_marker); |
| 265 |
TMPMEMFREE(counter); |
| 266 |
if (Paso_noError()) { |
| 267 |
if (verbose) { |
| 268 |
printf("RILU: %d unknowns eliminated. %d left.\n",out->n_F,n-out->n_F); |
| 269 |
if (out->n_C>0) { |
| 270 |
printf("timing: RILU: MIS/reordering/elemination : %e/%e/%e\n",time2,time0,time1); |
| 271 |
} else { |
| 272 |
printf("timing: RILU: MIS: %e\n",time2); |
| 273 |
} |
| 274 |
} |
| 275 |
return out; |
| 276 |
} else { |
| 277 |
Paso_Solver_RILU_free(out); |
| 278 |
return NULL; |
| 279 |
} |
| 280 |
} |
| 281 |
|
| 282 |
/**************************************************************/ |
| 283 |
|
| 284 |
/* apply RILU precondition b-> x |
| 285 |
|
| 286 |
in fact it solves |
| 287 |
|
| 288 |
[ I 0 ] [ A_FF 0 ] [ I invA_FF*A_FF ] [ x_F ] = [b_F] |
| 289 |
[ A_CF*invA_FF I ] [ 0 S ] [ 0 I ] [ x_C ] = [b_C] |
| 290 |
|
| 291 |
in the form |
| 292 |
|
| 293 |
b->[b_F,b_C] |
| 294 |
x_F=invA_FF*b_F |
| 295 |
b_C=b_C-A_CF*x_F |
| 296 |
x_C=RILU(b_C) |
| 297 |
b_F=b_F-A_FC*x_C |
| 298 |
x_F=invA_FF*b_F |
| 299 |
x<-[x_F,x_C] |
| 300 |
|
| 301 |
should be called within a parallel region |
| 302 |
barrier synconization should be performed to make sure that the input vector available |
| 303 |
|
| 304 |
*/ |
| 305 |
|
| 306 |
void Paso_Solver_solveRILU(Paso_Solver_RILU * rilu, double * x, double * b) { |
| 307 |
dim_t i,k; |
| 308 |
dim_t n_block=rilu->n_block; |
| 309 |
|
| 310 |
if (rilu->n_C==0) { |
| 311 |
/* x=invA_FF*b */ |
| 312 |
Paso_Solver_applyBlockDiagonalMatrix(n_block,rilu->n_F,rilu->inv_A_FF,rilu->A_FF_pivot,x,b); |
| 313 |
} else { |
| 314 |
/* b->[b_F,b_C] */ |
| 315 |
if (n_block==1) { |
| 316 |
#pragma omp for private(i) schedule(static) |
| 317 |
for (i=0;i<rilu->n_F;++i) rilu->b_F[i]=b[rilu->rows_in_F[i]]; |
| 318 |
#pragma omp for private(i) schedule(static) |
| 319 |
for (i=0;i<rilu->n_C;++i) rilu->b_C[i]=b[rilu->rows_in_C[i]]; |
| 320 |
} else { |
| 321 |
#pragma omp for private(i,k) schedule(static) |
| 322 |
for (i=0;i<rilu->n_F;++i) |
| 323 |
for (k=0;k<n_block;k++) rilu->b_F[rilu->n_block*i+k]=b[n_block*rilu->rows_in_F[i]+k]; |
| 324 |
#pragma omp for private(i,k) schedule(static) |
| 325 |
for (i=0;i<rilu->n_C;++i) |
| 326 |
for (k=0;k<n_block;k++) rilu->b_C[rilu->n_block*i+k]=b[n_block*rilu->rows_in_C[i]+k]; |
| 327 |
} |
| 328 |
/* x_F=invA_FF*b_F */ |
| 329 |
Paso_Solver_applyBlockDiagonalMatrix(n_block,rilu->n_F,rilu->inv_A_FF,rilu->A_FF_pivot,rilu->x_F,rilu->b_F); |
| 330 |
/* b_C=b_C-A_CF*x_F */ |
| 331 |
Paso_SystemMatrix_MatrixVector_CSR_OFFSET0(-1.,rilu->A_CF,rilu->x_F,1.,rilu->b_C); |
| 332 |
/* x_C=RILU(b_C) */ |
| 333 |
Paso_Solver_solveRILU(rilu->RILU_of_Schur,rilu->x_C,rilu->b_C); |
| 334 |
/* b_F=b_F-A_FC*x_C */ |
| 335 |
Paso_SystemMatrix_MatrixVector_CSR_OFFSET0(-1.,rilu->A_FC,rilu->x_C,1.,rilu->b_F); |
| 336 |
/* x_F=invA_FF*b_F */ |
| 337 |
Paso_Solver_applyBlockDiagonalMatrix(n_block,rilu->n_F,rilu->inv_A_FF,rilu->A_FF_pivot,rilu->x_F,rilu->b_F); |
| 338 |
/* x<-[x_F,x_C] */ |
| 339 |
if (n_block==1) { |
| 340 |
#pragma omp for private(i) schedule(static) |
| 341 |
for (i=0;i<rilu->n;++i) { |
| 342 |
if (rilu->mask_C[i]>-1) { |
| 343 |
x[i]=rilu->x_C[rilu->mask_C[i]]; |
| 344 |
} else { |
| 345 |
x[i]=rilu->x_F[rilu->mask_F[i]]; |
| 346 |
} |
| 347 |
} |
| 348 |
} else { |
| 349 |
#pragma omp for private(i,k) schedule(static) |
| 350 |
for (i=0;i<rilu->n;++i) { |
| 351 |
if (rilu->mask_C[i]>-1) { |
| 352 |
for (k=0;k<n_block;k++) x[n_block*i+k]=rilu->x_C[n_block*rilu->mask_C[i]+k]; |
| 353 |
} else { |
| 354 |
for (k=0;k<n_block;k++) x[n_block*i+k]=rilu->x_F[n_block*rilu->mask_F[i]+k]; |
| 355 |
} |
| 356 |
} |
| 357 |
} |
| 358 |
/* all done */ |
| 359 |
} |
| 360 |
return; |
| 361 |
} |
| 362 |
|
| 363 |
/* |
| 364 |
* $Log$ |
| 365 |
* |
| 366 |
*/ |
Parent Directory
| 
Revision Log