1 /* 2 Defines projective product routines where A is a SeqAIJ matrix 3 C = P^T * A * P 4 */ 5 6 #include "src/mat/impls/aij/seq/aij.h" /*I "petscmat.h" I*/ 7 #include "src/mat/utils/freespace.h" 8 #include "petscbt.h" 9 10 #undef __FUNCT__ 11 #define __FUNCT__ "MatPtAP_SeqAIJ_SeqAIJ" 12 PetscErrorCode MatPtAP_SeqAIJ(Mat A,Mat P,MatReuse scall,PetscReal fill,Mat *C) 13 { 14 PetscErrorCode ierr; 15 16 PetscFunctionBegin; 17 if (scall == MAT_INITIAL_MATRIX){ 18 ierr = PetscLogEventBegin(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 19 ierr = MatPtAPSymbolic(A,P,fill,C);CHKERRQ(ierr); 20 ierr = PetscLogEventEnd(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr); 21 } 22 ierr = PetscLogEventBegin(MAT_PtAPNumeric,A,P,0,0);CHKERRQ(ierr); 23 ierr = MatPtAPNumeric(A,P,*C);CHKERRQ(ierr); 24 ierr = PetscLogEventEnd(MAT_PtAPNumeric,A,P,0,0);CHKERRQ(ierr); 25 PetscFunctionReturn(0); 26 } 27 28 #undef __FUNCT__ 29 #define __FUNCT__ "MatPtAPSymbolic_SeqAIJ" 30 PetscErrorCode MatPtAPSymbolic_SeqAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 31 { 32 PetscErrorCode ierr; 33 34 PetscFunctionBegin; 35 if (!P->ops->ptapsymbolic_seqaij) { 36 SETERRQ2(PETSC_ERR_SUP,"Not implemented for A=%s and P=%s",A->type_name,P->type_name); 37 } 38 ierr = (*P->ops->ptapsymbolic_seqaij)(A,P,fill,C);CHKERRQ(ierr); 39 PetscFunctionReturn(0); 40 } 41 42 #undef __FUNCT__ 43 #define __FUNCT__ "MatPtAPNumeric_SeqAIJ" 44 PetscErrorCode MatPtAPNumeric_SeqAIJ(Mat A,Mat P,Mat C) 45 { 46 PetscErrorCode ierr; 47 48 PetscFunctionBegin; 49 if (!P->ops->ptapnumeric_seqaij) { 50 SETERRQ2(PETSC_ERR_SUP,"Not implemented for A=%s and P=%s",A->type_name,P->type_name); 51 } 52 ierr = (*P->ops->ptapnumeric_seqaij)(A,P,C);CHKERRQ(ierr); 53 PetscFunctionReturn(0); 54 } 55 56 #undef __FUNCT__ 57 #define __FUNCT__ "MatPtAPSymbolic_SeqAIJ_SeqAIJ" 58 PetscErrorCode MatPtAPSymbolic_SeqAIJ_SeqAIJ(Mat A,Mat P,PetscReal fill,Mat *C) 59 { 60 PetscErrorCode ierr; 61 FreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL; 62 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*p = (Mat_SeqAIJ*)P->data,*c; 63 PetscInt *pti,*ptj,*ptJ,*ai=a->i,*aj=a->j,*ajj,*pi=p->i,*pj=p->j,*pjj; 64 PetscInt *ci,*cj,*ptadenserow,*ptasparserow,*ptaj; 65 PetscInt an=A->N,am=A->M,pn=P->N,pm=P->M; 66 PetscInt i,j,k,ptnzi,arow,anzj,ptanzi,prow,pnzj,cnzi,nlnk,*lnk; 67 MatScalar *ca; 68 PetscBT lnkbt; 69 70 PetscFunctionBegin; 71 /* Get ij structure of P^T */ 72 ierr = MatGetSymbolicTranspose_SeqAIJ(P,&pti,&ptj);CHKERRQ(ierr); 73 ptJ=ptj; 74 75 /* Allocate ci array, arrays for fill computation and */ 76 /* free space for accumulating nonzero column info */ 77 ierr = PetscMalloc((pn+1)*sizeof(PetscInt),&ci);CHKERRQ(ierr); 78 ci[0] = 0; 79 80 ierr = PetscMalloc((2*an+1)*sizeof(PetscInt),&ptadenserow);CHKERRQ(ierr); 81 ierr = PetscMemzero(ptadenserow,(2*an+1)*sizeof(PetscInt));CHKERRQ(ierr); 82 ptasparserow = ptadenserow + an; 83 84 /* create and initialize a linked list */ 85 nlnk = pn+1; 86 ierr = PetscLLCreate(pn,pn,nlnk,lnk,lnkbt);CHKERRQ(ierr); 87 88 /* Set initial free space to be nnz(A) scaled by aspect ratio of P. */ 89 /* This should be reasonable if sparsity of PtAP is similar to that of A. */ 90 ierr = GetMoreSpace((ai[am]/pm)*pn,&free_space); 91 current_space = free_space; 92 93 /* Determine symbolic info for each row of C: */ 94 for (i=0;i<pn;i++) { 95 ptnzi = pti[i+1] - pti[i]; 96 ptanzi = 0; 97 /* Determine symbolic row of PtA: */ 98 for (j=0;j<ptnzi;j++) { 99 arow = *ptJ++; 100 anzj = ai[arow+1] - ai[arow]; 101 ajj = aj + ai[arow]; 102 for (k=0;k<anzj;k++) { 103 if (!ptadenserow[ajj[k]]) { 104 ptadenserow[ajj[k]] = -1; 105 ptasparserow[ptanzi++] = ajj[k]; 106 } 107 } 108 } 109 /* Using symbolic info for row of PtA, determine symbolic info for row of C: */ 110 ptaj = ptasparserow; 111 cnzi = 0; 112 for (j=0;j<ptanzi;j++) { 113 prow = *ptaj++; 114 pnzj = pi[prow+1] - pi[prow]; 115 pjj = pj + pi[prow]; 116 /* add non-zero cols of P into the sorted linked list lnk */ 117 ierr = PetscLLAdd(pnzj,pjj,pn,nlnk,lnk,lnkbt);CHKERRQ(ierr); 118 cnzi += nlnk; 119 } 120 121 /* If free space is not available, make more free space */ 122 /* Double the amount of total space in the list */ 123 if (current_space->local_remaining<cnzi) { 124 ierr = GetMoreSpace(current_space->total_array_size,¤t_space);CHKERRQ(ierr); 125 } 126 127 /* Copy data into free space, and zero out denserows */ 128 ierr = PetscLLClean(pn,pn,cnzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 129 current_space->array += cnzi; 130 current_space->local_used += cnzi; 131 current_space->local_remaining -= cnzi; 132 133 for (j=0;j<ptanzi;j++) { 134 ptadenserow[ptasparserow[j]] = 0; 135 } 136 /* Aside: Perhaps we should save the pta info for the numerical factorization. */ 137 /* For now, we will recompute what is needed. */ 138 ci[i+1] = ci[i] + cnzi; 139 } 140 /* nnz is now stored in ci[ptm], column indices are in the list of free space */ 141 /* Allocate space for cj, initialize cj, and */ 142 /* destroy list of free space and other temporary array(s) */ 143 ierr = PetscMalloc((ci[pn]+1)*sizeof(PetscInt),&cj);CHKERRQ(ierr); 144 ierr = MakeSpaceContiguous(&free_space,cj);CHKERRQ(ierr); 145 ierr = PetscFree(ptadenserow);CHKERRQ(ierr); 146 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 147 148 /* Allocate space for ca */ 149 ierr = PetscMalloc((ci[pn]+1)*sizeof(MatScalar),&ca);CHKERRQ(ierr); 150 ierr = PetscMemzero(ca,(ci[pn]+1)*sizeof(MatScalar));CHKERRQ(ierr); 151 152 /* put together the new matrix */ 153 ierr = MatCreateSeqAIJWithArrays(A->comm,pn,pn,ci,cj,ca,C);CHKERRQ(ierr); 154 155 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 156 /* Since these are PETSc arrays, change flags to free them as necessary. */ 157 c = (Mat_SeqAIJ *)((*C)->data); 158 c->freedata = PETSC_TRUE; 159 c->nonew = 0; 160 161 /* Clean up. */ 162 ierr = MatRestoreSymbolicTranspose_SeqAIJ(P,&pti,&ptj);CHKERRQ(ierr); 163 164 PetscFunctionReturn(0); 165 } 166 167 #undef __FUNCT__ 168 #define __FUNCT__ "MatPtAPNumeric_SeqAIJ_SeqAIJ" 169 PetscErrorCode MatPtAPNumeric_SeqAIJ_SeqAIJ(Mat A,Mat P,Mat C) 170 { 171 PetscErrorCode ierr; 172 PetscInt flops=0; 173 Mat_SeqAIJ *a = (Mat_SeqAIJ *) A->data; 174 Mat_SeqAIJ *p = (Mat_SeqAIJ *) P->data; 175 Mat_SeqAIJ *c = (Mat_SeqAIJ *) C->data; 176 PetscInt *ai=a->i,*aj=a->j,*apj,*apjdense,*pi=p->i,*pj=p->j,*pJ=p->j,*pjj; 177 PetscInt *ci=c->i,*cj=c->j,*cjj; 178 PetscInt am=A->M,cn=C->N,cm=C->M; 179 PetscInt i,j,k,anzi,pnzi,apnzj,nextap,pnzj,prow,crow; 180 MatScalar *aa=a->a,*apa,*pa=p->a,*pA=p->a,*paj,*ca=c->a,*caj; 181 182 PetscFunctionBegin; 183 /* Allocate temporary array for storage of one row of A*P */ 184 ierr = PetscMalloc(cn*(sizeof(MatScalar)+2*sizeof(PetscInt)),&apa);CHKERRQ(ierr); 185 ierr = PetscMemzero(apa,cn*(sizeof(MatScalar)+2*sizeof(PetscInt)));CHKERRQ(ierr); 186 187 apj = (PetscInt *)(apa + cn); 188 apjdense = apj + cn; 189 190 /* Clear old values in C */ 191 ierr = PetscMemzero(ca,ci[cm]*sizeof(MatScalar));CHKERRQ(ierr); 192 193 for (i=0;i<am;i++) { 194 /* Form sparse row of A*P */ 195 anzi = ai[i+1] - ai[i]; 196 apnzj = 0; 197 for (j=0;j<anzi;j++) { 198 prow = *aj++; 199 pnzj = pi[prow+1] - pi[prow]; 200 pjj = pj + pi[prow]; 201 paj = pa + pi[prow]; 202 for (k=0;k<pnzj;k++) { 203 if (!apjdense[pjj[k]]) { 204 apjdense[pjj[k]] = -1; 205 apj[apnzj++] = pjj[k]; 206 } 207 apa[pjj[k]] += (*aa)*paj[k]; 208 } 209 flops += 2*pnzj; 210 aa++; 211 } 212 213 /* Sort the j index array for quick sparse axpy. */ 214 /* Note: a array does not need sorting as it is in dense storage locations. */ 215 ierr = PetscSortInt(apnzj,apj);CHKERRQ(ierr); 216 217 /* Compute P^T*A*P using outer product (P^T)[:,j]*(A*P)[j,:]. */ 218 pnzi = pi[i+1] - pi[i]; 219 for (j=0;j<pnzi;j++) { 220 nextap = 0; 221 crow = *pJ++; 222 cjj = cj + ci[crow]; 223 caj = ca + ci[crow]; 224 /* Perform sparse axpy operation. Note cjj includes apj. */ 225 for (k=0;nextap<apnzj;k++) { 226 if (cjj[k]==apj[nextap]) { 227 caj[k] += (*pA)*apa[apj[nextap++]]; 228 } 229 } 230 flops += 2*apnzj; 231 pA++; 232 } 233 234 /* Zero the current row info for A*P */ 235 for (j=0;j<apnzj;j++) { 236 apa[apj[j]] = 0.; 237 apjdense[apj[j]] = 0; 238 } 239 } 240 241 /* Assemble the final matrix and clean up */ 242 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 243 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 244 ierr = PetscFree(apa);CHKERRQ(ierr); 245 ierr = PetscLogFlops(flops);CHKERRQ(ierr); 246 247 PetscFunctionReturn(0); 248 } 249