Actual source code: matrart.c

petsc-3.7.3 2016-08-01
Report Typos and Errors
  2: /*
  3:   Defines projective product routines where A is a SeqAIJ matrix
  4:           C = R * A * R^T
  5: */

  7: #include <../src/mat/impls/aij/seq/aij.h>
  8: #include <../src/mat/utils/freespace.h>
  9: #include <../src/mat/impls/dense/seq/dense.h> /*I "petscmat.h" I*/

 13: PetscErrorCode MatDestroy_SeqAIJ_RARt(Mat A)
 14: {
 16:   Mat_SeqAIJ     *a    = (Mat_SeqAIJ*)A->data;
 17:   Mat_RARt       *rart = a->rart;

 20:   MatTransposeColoringDestroy(&rart->matcoloring);
 21:   MatDestroy(&rart->Rt);
 22:   MatDestroy(&rart->RARt);
 23:   MatDestroy(&rart->ARt);
 24:   PetscFree(rart->work);

 26:   A->ops->destroy = rart->destroy;
 27:   if (A->ops->destroy) {
 28:     (*A->ops->destroy)(A);
 29:   }
 30:   PetscFree(rart);
 31:   return(0);
 32: }

 36: PetscErrorCode MatRARtSymbolic_SeqAIJ_SeqAIJ_colorrart(Mat A,Mat R,PetscReal fill,Mat *C)
 37: {
 38:   PetscErrorCode       ierr;
 39:   Mat                  P;
 40:   PetscInt             *rti,*rtj;
 41:   Mat_RARt             *rart;
 42:   MatColoring          coloring;
 43:   MatTransposeColoring matcoloring;
 44:   ISColoring           iscoloring;
 45:   Mat                  Rt_dense,RARt_dense;
 46:   Mat_SeqAIJ           *c;

 49:   /* create symbolic P=Rt */
 50:   MatGetSymbolicTranspose_SeqAIJ(R,&rti,&rtj);
 51:   MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,R->cmap->n,R->rmap->n,rti,rtj,NULL,&P);

 53:   /* get symbolic C=Pt*A*P */
 54:   MatPtAPSymbolic_SeqAIJ_SeqAIJ_SparseAxpy(A,P,fill,C);
 55:   MatSetBlockSizes(*C,PetscAbs(R->rmap->bs),PetscAbs(R->rmap->bs));
 56:   (*C)->ops->rartnumeric = MatRARtNumeric_SeqAIJ_SeqAIJ_colorrart;

 58:   /* create a supporting struct */
 59:   PetscNew(&rart);
 60:   c       = (Mat_SeqAIJ*)(*C)->data;
 61:   c->rart = rart;

 63:   /* ------ Use coloring ---------- */
 64:   /* inode causes memory problem, don't know why */
 65:   if (c->inode.use) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MAT_USE_INODES is not supported. Use '-mat_no_inode'");

 67:   /* Create MatTransposeColoring from symbolic C=R*A*R^T */
 68:   MatColoringCreate(*C,&coloring);
 69:   MatColoringSetDistance(coloring,2);
 70:   MatColoringSetType(coloring,MATCOLORINGSL);
 71:   MatColoringSetFromOptions(coloring);
 72:   MatColoringApply(coloring,&iscoloring);
 73:   MatColoringDestroy(&coloring);
 74:   MatTransposeColoringCreate(*C,iscoloring,&matcoloring);

 76:   rart->matcoloring = matcoloring;
 77:   ISColoringDestroy(&iscoloring);

 79:   /* Create Rt_dense */
 80:   MatCreate(PETSC_COMM_SELF,&Rt_dense);
 81:   MatSetSizes(Rt_dense,A->cmap->n,matcoloring->ncolors,A->cmap->n,matcoloring->ncolors);
 82:   MatSetType(Rt_dense,MATSEQDENSE);
 83:   MatSeqDenseSetPreallocation(Rt_dense,NULL);

 85:   Rt_dense->assembled = PETSC_TRUE;
 86:   rart->Rt            = Rt_dense;

 88:   /* Create RARt_dense = R*A*Rt_dense */
 89:   MatCreate(PETSC_COMM_SELF,&RARt_dense);
 90:   MatSetSizes(RARt_dense,(*C)->rmap->n,matcoloring->ncolors,(*C)->rmap->n,matcoloring->ncolors);
 91:   MatSetType(RARt_dense,MATSEQDENSE);
 92:   MatSeqDenseSetPreallocation(RARt_dense,NULL);

 94:   rart->RARt = RARt_dense;

 96:   /* Allocate work array to store columns of A*R^T used in MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqDense() */
 97:   PetscMalloc1(A->rmap->n*4,&rart->work);

 99:   rart->destroy      = (*C)->ops->destroy;
100:   (*C)->ops->destroy = MatDestroy_SeqAIJ_RARt;

102:   /* clean up */
103:   MatRestoreSymbolicTranspose_SeqAIJ(R,&rti,&rtj);
104:   MatDestroy(&P);

106: #if defined(PETSC_USE_INFO)
107:   {
108:     PetscReal density= (PetscReal)(c->nz)/(RARt_dense->rmap->n*RARt_dense->cmap->n);
109:     PetscInfo(*C,"C=R*(A*Rt) via coloring C - use sparse-dense inner products\n");
110:     PetscInfo6(*C,"RARt_den %D %D; Rt %D %D (RARt->nz %D)/(m*ncolors)=%g\n",RARt_dense->rmap->n,RARt_dense->cmap->n,R->cmap->n,R->rmap->n,c->nz,density);
111:   }
112: #endif
113:   return(0);
114: }

116: /*
117:  RAB = R * A * B, R and A in seqaij format, B in dense format;
118: */
121: PetscErrorCode MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqDense(Mat R,Mat A,Mat B,Mat RAB,PetscScalar *work)
122: {
123:   Mat_SeqAIJ     *a=(Mat_SeqAIJ*)A->data,*r=(Mat_SeqAIJ*)R->data;
125:   PetscScalar    *b,r1,r2,r3,r4,*b1,*b2,*b3,*b4;
126:   MatScalar      *aa,*ra;
127:   PetscInt       cn =B->cmap->n,bm=B->rmap->n,col,i,j,n,*ai=a->i,*aj,am=A->rmap->n;
128:   PetscInt       am2=2*am,am3=3*am,bm4=4*bm;
129:   PetscScalar    *d,*c,*c2,*c3,*c4;
130:   PetscInt       *rj,rm=R->rmap->n,dm=RAB->rmap->n,dn=RAB->cmap->n;
131:   PetscInt       rm2=2*rm,rm3=3*rm,colrm;

134:   if (!dm || !dn) return(0);
135:   if (bm != A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Number columns in A %D not equal rows in B %D\n",A->cmap->n,bm);
136:   if (am != R->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Number columns in R %D not equal rows in A %D\n",R->cmap->n,am);
137:   if (R->rmap->n != RAB->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Number rows in RAB %D not equal rows in R %D\n",RAB->rmap->n,R->rmap->n);
138:   if (B->cmap->n != RAB->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Number columns in RAB %D not equal columns in B %D\n",RAB->cmap->n,B->cmap->n);

140:   { /* 
141:      This approach is not as good as original ones (will be removed later), but it reveals that
142:      AB_den=A*B takes almost all execution time in R*A*B for src/ksp/ksp/examples/tutorials/ex56.c
143:      */
144:     PetscBool via_matmatmult=PETSC_FALSE;
145:     PetscOptionsGetBool(NULL,NULL,"-matrart_via_matmatmult",&via_matmatmult,NULL);
146:     if (via_matmatmult) {
147:       Mat AB_den;
148:       MatMatMultSymbolic_SeqAIJ_SeqDense(A,B,0.0,&AB_den);
149:       MatMatMultNumeric_SeqAIJ_SeqDense(A,B,AB_den);
150:       MatMatMultNumeric_SeqAIJ_SeqDense(R,AB_den,RAB);
151:       MatDestroy(&AB_den);
152:       return(0);
153:     }
154:   }

156:   MatDenseGetArray(B,&b);
157:   MatDenseGetArray(RAB,&d);
158:   b1   = b; b2 = b1 + bm; b3 = b2 + bm; b4 = b3 + bm;
159:   c    = work; c2 = c + am; c3 = c2 + am; c4 = c3 + am;
160:   for (col=0; col<cn-4; col += 4) {  /* over columns of C */
161:     for (i=0; i<am; i++) {        /* over rows of A in those columns */
162:       r1 = r2 = r3 = r4 = 0.0;
163:       n  = ai[i+1] - ai[i];
164:       aj = a->j + ai[i];
165:       aa = a->a + ai[i];
166:       for (j=0; j<n; j++) {
167:         r1 += (*aa)*b1[*aj];
168:         r2 += (*aa)*b2[*aj];
169:         r3 += (*aa)*b3[*aj];
170:         r4 += (*aa++)*b4[*aj++];
171:       }
172:       c[i]       = r1;
173:       c[am  + i] = r2;
174:       c[am2 + i] = r3;
175:       c[am3 + i] = r4;
176:     }
177:     b1 += bm4;
178:     b2 += bm4;
179:     b3 += bm4;
180:     b4 += bm4;

182:     /* RAB[:,col] = R*C[:,col] */
183:     colrm = col*rm;
184:     for (i=0; i<rm; i++) {        /* over rows of R in those columns */
185:       r1 = r2 = r3 = r4 = 0.0;
186:       n  = r->i[i+1] - r->i[i];
187:       rj = r->j + r->i[i];
188:       ra = r->a + r->i[i];
189:       for (j=0; j<n; j++) {
190:         r1 += (*ra)*c[*rj];
191:         r2 += (*ra)*c2[*rj];
192:         r3 += (*ra)*c3[*rj];
193:         r4 += (*ra++)*c4[*rj++];
194:       }
195:       d[colrm + i]       = r1;
196:       d[colrm + rm + i]  = r2;
197:       d[colrm + rm2 + i] = r3;
198:       d[colrm + rm3 + i] = r4;
199:     }
200:   }
201:   for (; col<cn; col++) {     /* over extra columns of C */
202:     for (i=0; i<am; i++) {  /* over rows of A in those columns */
203:       r1 = 0.0;
204:       n  = a->i[i+1] - a->i[i];
205:       aj = a->j + a->i[i];
206:       aa = a->a + a->i[i];
207:       for (j=0; j<n; j++) {
208:         r1 += (*aa++)*b1[*aj++];
209:       }
210:       c[i] = r1;
211:     }
212:     b1 += bm;

214:     for (i=0; i<rm; i++) {  /* over rows of R in those columns */
215:       r1 = 0.0;
216:       n  = r->i[i+1] - r->i[i];
217:       rj = r->j + r->i[i];
218:       ra = r->a + r->i[i];
219:       for (j=0; j<n; j++) {
220:         r1 += (*ra++)*c[*rj++];
221:       }
222:       d[col*rm + i] = r1;
223:     }
224:   }
225:   PetscLogFlops(cn*2.0*(a->nz + r->nz));

227:   MatDenseRestoreArray(B,&b);
228:   MatDenseRestoreArray(RAB,&d);
229:   MatAssemblyBegin(RAB,MAT_FINAL_ASSEMBLY);
230:   MatAssemblyEnd(RAB,MAT_FINAL_ASSEMBLY);
231:   return(0);
232: }

236: PetscErrorCode MatRARtNumeric_SeqAIJ_SeqAIJ_colorrart(Mat A,Mat R,Mat C)
237: {
238:   PetscErrorCode       ierr;
239:   Mat_SeqAIJ           *c = (Mat_SeqAIJ*)C->data;
240:   Mat_RARt             *rart=c->rart;
241:   MatTransposeColoring matcoloring;
242:   Mat                  Rt,RARt;

245:   /* Get dense Rt by Apply MatTransposeColoring to R */
246:   matcoloring = rart->matcoloring;
247:   Rt          = rart->Rt;
248:   MatTransColoringApplySpToDen(matcoloring,R,Rt);

250:   /* Get dense RARt = R*A*Rt -- dominates! */
251:   RARt = rart->RARt;
252:   MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqDense(R,A,Rt,RARt,rart->work);

254:   /* Recover C from C_dense */
255:   MatTransColoringApplyDenToSp(matcoloring,RARt,C);
256:   return(0);
257: }

261: PetscErrorCode MatRARtSymbolic_SeqAIJ_SeqAIJ_matmattransposemult(Mat A,Mat R,PetscReal fill,Mat *C)
262: {
263:   PetscErrorCode  ierr;
264:   Mat             ARt,RARt;
265:   Mat_SeqAIJ     *c;
266:   Mat_RARt       *rart;

269:   /* must use '-mat_no_inode' with '-matmattransmult_color 1' - do not knwo why? */
270:   MatMatTransposeMultSymbolic_SeqAIJ_SeqAIJ(A,R,fill,&ARt);
271:   MatMatMultSymbolic_SeqAIJ_SeqAIJ(R,ARt,fill,&RARt);
272:   *C                     = RARt;
273:   RARt->ops->rartnumeric = MatRARtNumeric_SeqAIJ_SeqAIJ_matmattransposemult;

275:   PetscNew(&rart);
276:   c         = (Mat_SeqAIJ*)(*C)->data;
277:   c->rart   = rart;
278:   rart->ARt = ARt;
279:   rart->destroy      = RARt->ops->destroy;
280:   RARt->ops->destroy = MatDestroy_SeqAIJ_RARt;
281: #if defined(PETSC_USE_INFO)
282:   PetscInfo(*C,"Use ARt=A*R^T, C=R*ARt via MatMatTransposeMult(). Coloring can be applied to A*R^T.\n");
283: #endif
284:   return(0);
285: }

289: PetscErrorCode MatRARtNumeric_SeqAIJ_SeqAIJ_matmattransposemult(Mat A,Mat R,Mat C)
290: {
291:   PetscErrorCode  ierr;
292:   Mat_SeqAIJ      *c=(Mat_SeqAIJ*)C->data;
293:   Mat_RARt        *rart=c->rart;
294:   Mat             ARt=rart->ARt;
295: 
297:   MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ(A,R,ARt); /* dominate! */
298:   MatMatMultNumeric_SeqAIJ_SeqAIJ(R,ARt,C);
299:   return(0);
300: }

304: PetscErrorCode MatRARtSymbolic_SeqAIJ_SeqAIJ(Mat A,Mat R,PetscReal fill,Mat *C)
305: {
306:   PetscErrorCode  ierr;
307:   Mat             Rt;
308:   Mat_SeqAIJ      *c;
309:   Mat_RARt        *rart;

312:   MatTranspose_SeqAIJ(R,MAT_INITIAL_MATRIX,&Rt);
313:   MatMatMatMultSymbolic_SeqAIJ_SeqAIJ_SeqAIJ(R,A,Rt,fill,C);

315:   PetscNew(&rart);
316:   rart->Rt = Rt;
317:   c        = (Mat_SeqAIJ*)(*C)->data;
318:   c->rart  = rart;
319:   rart->destroy          = (*C)->ops->destroy;
320:   (*C)->ops->destroy     = MatDestroy_SeqAIJ_RARt;
321:   (*C)->ops->rartnumeric = MatRARtNumeric_SeqAIJ_SeqAIJ;
322: #if defined(PETSC_USE_INFO)
323:   PetscInfo(*C,"Use Rt=R^T and C=R*A*Rt via MatMatMatMult() to avoid sparse inner products\n");
324: #endif
325:   return(0);
326: }

330: PetscErrorCode MatRARtNumeric_SeqAIJ_SeqAIJ(Mat A,Mat R,Mat C)
331: {
332:   PetscErrorCode  ierr;
333:   Mat_SeqAIJ      *c = (Mat_SeqAIJ*)C->data;
334:   Mat_RARt        *rart = c->rart;
335:   Mat             Rt = rart->Rt;

338:   MatTranspose_SeqAIJ(R,MAT_REUSE_MATRIX,&Rt);
339:   MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ(R,A,Rt,C);
340:   return(0);
341: }

345: PetscErrorCode MatRARt_SeqAIJ_SeqAIJ(Mat A,Mat R,MatReuse scall,PetscReal fill,Mat *C)
346: {
348:   const char     *algTypes[3] = {"matmatmatmult","matmattransposemult","coloring_rart"};
349:   PetscInt       alg=0; /* set default algorithm */
350: 
352:   if (scall == MAT_INITIAL_MATRIX) {
353:     PetscObjectOptionsBegin((PetscObject)A);
354:     PetscOptionsEList("-matrart_via","Algorithmic approach","MatRARt",algTypes,3,algTypes[0],&alg,NULL);
355:     PetscOptionsEnd();

357:     PetscLogEventBegin(MAT_RARtSymbolic,A,R,0,0);
358:     switch (alg) {
359:     case 1:
360:       /* via matmattransposemult: ARt=A*R^T, C=R*ARt - matrix coloring can be applied to A*R^T */
361:       MatRARtSymbolic_SeqAIJ_SeqAIJ_matmattransposemult(A,R,fill,C);
362:       break;
363:     case 2:
364:       /* via coloring_rart: apply coloring C = R*A*R^T                          */
365:       MatRARtSymbolic_SeqAIJ_SeqAIJ_colorrart(A,R,fill,C);
366:       break;
367:     default:
368:       /* via matmatmatmult: Rt=R^T, C=R*A*Rt - avoid inefficient sparse inner products */
369:       MatRARtSymbolic_SeqAIJ_SeqAIJ(A,R,fill,C);
370:       break;
371:     }
372:     PetscLogEventEnd(MAT_RARtSymbolic,A,R,0,0);
373:   }

375:   PetscLogEventBegin(MAT_RARtNumeric,A,R,0,0);
376:   (*(*C)->ops->rartnumeric)(A,R,*C);
377:   PetscLogEventEnd(MAT_RARtNumeric,A,R,0,0);
378:   return(0);
379: }