1 2 #if !defined(__AIJ_H) 3 #define __AIJ_H 4 5 #include "private/matimpl.h" 6 7 /* 8 Struct header shared by SeqAIJ, SeqBAIJ and SeqSBAIJ matrix formats 9 */ 10 #define SEQAIJHEADER(datatype) \ 11 PetscBool roworiented; /* if true, row-oriented input, default */\ 12 PetscInt nonew; /* 1 don't add new nonzeros, -1 generate error on new */\ 13 PetscInt nounused; /* -1 generate error on unused space */\ 14 PetscBool singlemalloc; /* if true a, i, and j have been obtained with one big malloc */\ 15 PetscInt maxnz; /* allocated nonzeros */\ 16 PetscInt *imax; /* maximum space allocated for each row */\ 17 PetscInt *ilen; /* actual length of each row */\ 18 PetscBool free_imax_ilen; \ 19 PetscInt reallocs; /* number of mallocs done during MatSetValues() \ 20 as more values are set than were prealloced */\ 21 PetscInt rmax; /* max nonzeros in any row */\ 22 PetscBool keepnonzeropattern; /* keeps matrix structure same in calls to MatZeroRows()*/\ 23 PetscBool ignorezeroentries; \ 24 PetscInt *xtoy,*xtoyB; /* map nonzero pattern of X into Y's, used by MatAXPY() */\ 25 Mat XtoY; /* used by MatAXPY() */\ 26 PetscBool free_ij; /* free the column indices j and row offsets i when the matrix is destroyed */ \ 27 PetscBool free_a; /* free the numerical values when matrix is destroy */ \ 28 Mat_CompressedRow compressedrow; /* use compressed row format */ \ 29 PetscInt nz; /* nonzeros */ \ 30 PetscInt *i; /* pointer to beginning of each row */ \ 31 PetscInt *j; /* column values: j + i[k] - 1 is start of row k */ \ 32 PetscInt *diag; /* pointers to diagonal elements */ \ 33 PetscBool free_diag; \ 34 datatype *a; /* nonzero elements */ \ 35 PetscScalar *solve_work; /* work space used in MatSolve */ \ 36 IS row, col, icol; /* index sets, used for reorderings */ \ 37 PetscBool pivotinblocks; /* pivot inside factorization of each diagonal block */ \ 38 Mat parent /* set if this matrix was formed with MatDuplicate(...,MAT_SHARE_NONZERO_PATTERN,....); 39 means that this shares some data structures with the parent including diag, ilen, imax, i, j */ 40 41 /* 42 MATSEQAIJ format - Compressed row storage (also called Yale sparse matrix 43 format) or compressed sparse row (CSR). The i[] and j[] arrays start at 0. For example, 44 j[i[k]+p] is the pth column in row k. Note that the diagonal 45 matrix elements are stored with the rest of the nonzeros (not separately). 46 */ 47 48 /* Info about i-nodes (identical nodes) helper class for SeqAIJ */ 49 typedef struct { 50 MatScalar *bdiag,*ibdiag,*ssor_work; /* diagonal blocks of matrix used for MatSOR_SeqAIJ_Inode() */ 51 PetscInt bdiagsize; /* length of bdiag and ibdiag */ 52 PetscBool ibdiagvalid; /* do ibdiag[] and bdiag[] contain the most recent values */ 53 54 PetscBool use; 55 PetscInt node_count; /* number of inodes */ 56 PetscInt *size; /* size of each inode */ 57 PetscInt limit; /* inode limit */ 58 PetscInt max_limit; /* maximum supported inode limit */ 59 PetscBool checked; /* if inodes have been checked for */ 60 } Mat_SeqAIJ_Inode; 61 62 EXTERN PetscErrorCode MatView_SeqAIJ_Inode(Mat,PetscViewer); 63 EXTERN PetscErrorCode MatAssemblyEnd_SeqAIJ_Inode(Mat,MatAssemblyType); 64 EXTERN PetscErrorCode MatDestroy_SeqAIJ_Inode(Mat); 65 EXTERN PetscErrorCode MatCreate_SeqAIJ_Inode(Mat); 66 EXTERN PetscErrorCode MatSetOption_SeqAIJ_Inode(Mat,MatOption,PetscBool ); 67 EXTERN PetscErrorCode MatDuplicate_SeqAIJ_Inode(Mat,MatDuplicateOption,Mat*); 68 EXTERN PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat,Mat,MatDuplicateOption,PetscBool ); 69 EXTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_Inode_inplace(Mat,Mat,const MatFactorInfo*); 70 EXTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_Inode(Mat,Mat,const MatFactorInfo*); 71 72 typedef struct { 73 SEQAIJHEADER(MatScalar); 74 Mat_SeqAIJ_Inode inode; 75 MatScalar *saved_values; /* location for stashing nonzero values of matrix */ 76 77 PetscScalar *idiag,*mdiag,*ssor_work; /* inverse of diagonal entries, diagonal values and workspace for Eisenstat trick */ 78 PetscBool idiagvalid; /* current idiag[] and mdiag[] are valid */ 79 PetscScalar fshift,omega; /* last used omega and fshift */ 80 81 ISColoring coloring; /* set with MatADSetColoring() used by MatADSetValues() */ 82 } Mat_SeqAIJ; 83 84 /* 85 Frees the a, i, and j arrays from the XAIJ (AIJ, BAIJ, and SBAIJ) matrix types 86 */ 87 #undef __FUNCT__ 88 #define __FUNCT__ "MatSeqXAIJFreeAIJ" 89 PETSC_STATIC_INLINE PetscErrorCode MatSeqXAIJFreeAIJ(Mat AA,MatScalar **a,PetscInt **j,PetscInt **i) 90 { 91 PetscErrorCode ierr; 92 Mat_SeqAIJ *A = (Mat_SeqAIJ*) AA->data; 93 if (A->singlemalloc) { 94 ierr = PetscFree3(*a,*j,*i);CHKERRQ(ierr); 95 } else { 96 if (A->free_a && *a) {ierr = PetscFree(*a);CHKERRQ(ierr);} 97 if (A->free_ij && *j) {ierr = PetscFree(*j);CHKERRQ(ierr);} 98 if (A->free_ij && *i) {ierr = PetscFree(*i);CHKERRQ(ierr);} 99 } 100 *a = 0; *j = 0; *i = 0; 101 return 0; 102 } 103 /* 104 Allocates larger a, i, and j arrays for the XAIJ (AIJ, BAIJ, and SBAIJ) matrix types 105 This is a macro because it takes the datatype as an argument which can be either a Mat or a MatScalar 106 */ 107 #define MatSeqXAIJReallocateAIJ(Amat,AM,BS2,NROW,ROW,COL,RMAX,AA,AI,AJ,RP,AP,AIMAX,NONEW,datatype) \ 108 if (NROW >= RMAX) {\ 109 Mat_SeqAIJ *Ain = (Mat_SeqAIJ*)Amat->data;\ 110 /* there is no extra room in row, therefore enlarge */ \ 111 PetscInt CHUNKSIZE = 15,new_nz = AI[AM] + CHUNKSIZE,len,*new_i=0,*new_j=0; \ 112 datatype *new_a; \ 113 \ 114 if (NONEW == -2) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"New nonzero at (%D,%D) caused a malloc",ROW,COL); \ 115 /* malloc new storage space */ \ 116 ierr = PetscMalloc3(BS2*new_nz,datatype,&new_a,new_nz,PetscInt,&new_j,AM+1,PetscInt,&new_i);CHKERRQ(ierr);\ 117 \ 118 /* copy over old data into new slots */ \ 119 for (ii=0; ii<ROW+1; ii++) {new_i[ii] = AI[ii];} \ 120 for (ii=ROW+1; ii<AM+1; ii++) {new_i[ii] = AI[ii]+CHUNKSIZE;} \ 121 ierr = PetscMemcpy(new_j,AJ,(AI[ROW]+NROW)*sizeof(PetscInt));CHKERRQ(ierr); \ 122 len = (new_nz - CHUNKSIZE - AI[ROW] - NROW); \ 123 ierr = PetscMemcpy(new_j+AI[ROW]+NROW+CHUNKSIZE,AJ+AI[ROW]+NROW,len*sizeof(PetscInt));CHKERRQ(ierr); \ 124 ierr = PetscMemcpy(new_a,AA,BS2*(AI[ROW]+NROW)*sizeof(datatype));CHKERRQ(ierr); \ 125 ierr = PetscMemzero(new_a+BS2*(AI[ROW]+NROW),BS2*CHUNKSIZE*sizeof(datatype));CHKERRQ(ierr);\ 126 ierr = PetscMemcpy(new_a+BS2*(AI[ROW]+NROW+CHUNKSIZE),AA+BS2*(AI[ROW]+NROW),BS2*len*sizeof(datatype));CHKERRQ(ierr); \ 127 /* free up old matrix storage */ \ 128 ierr = MatSeqXAIJFreeAIJ(A,&Ain->a,&Ain->j,&Ain->i);CHKERRQ(ierr);\ 129 AA = new_a; \ 130 Ain->a = (MatScalar*) new_a; \ 131 AI = Ain->i = new_i; AJ = Ain->j = new_j; \ 132 Ain->singlemalloc = PETSC_TRUE; \ 133 \ 134 RP = AJ + AI[ROW]; AP = AA + BS2*AI[ROW]; \ 135 RMAX = AIMAX[ROW] = AIMAX[ROW] + CHUNKSIZE; \ 136 Ain->maxnz += BS2*CHUNKSIZE; \ 137 Ain->reallocs++; \ 138 } \ 139 140 141 EXTERN_C_BEGIN 142 EXTERN PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat,PetscInt,const PetscInt*); 143 EXTERN_C_END 144 EXTERN PetscErrorCode MatILUFactorSymbolic_SeqAIJ_inplace(Mat,Mat,IS,IS,const MatFactorInfo*); 145 EXTERN PetscErrorCode MatILUFactorSymbolic_SeqAIJ(Mat,Mat,IS,IS,const MatFactorInfo*); 146 EXTERN PetscErrorCode MatILUFactorSymbolic_SeqAIJ_ilu0(Mat,Mat,IS,IS,const MatFactorInfo*); 147 148 EXTERN PetscErrorCode MatICCFactorSymbolic_SeqAIJ_inplace(Mat,Mat,IS,const MatFactorInfo*); 149 EXTERN PetscErrorCode MatICCFactorSymbolic_SeqAIJ(Mat,Mat,IS,const MatFactorInfo*); 150 EXTERN PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJ_inplace(Mat,Mat,IS,const MatFactorInfo*); 151 EXTERN PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJ(Mat,Mat,IS,const MatFactorInfo*); 152 EXTERN PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ_inplace(Mat,Mat,const MatFactorInfo*); 153 EXTERN PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ(Mat,Mat,const MatFactorInfo*); 154 EXTERN PetscErrorCode MatDuplicate_SeqAIJ(Mat,MatDuplicateOption,Mat*); 155 EXTERN PetscErrorCode MatCopy_SeqAIJ(Mat,Mat,MatStructure); 156 EXTERN PetscErrorCode MatMissingDiagonal_SeqAIJ(Mat,PetscBool *,PetscInt*); 157 EXTERN PetscErrorCode MatMarkDiagonal_SeqAIJ(Mat); 158 159 EXTERN PetscErrorCode MatMult_SeqAIJ(Mat A,Vec,Vec); 160 EXTERN PetscErrorCode MatMultAdd_SeqAIJ(Mat A,Vec,Vec,Vec); 161 EXTERN PetscErrorCode MatMultTranspose_SeqAIJ(Mat A,Vec,Vec); 162 EXTERN PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat A,Vec,Vec,Vec); 163 EXTERN PetscErrorCode MatSOR_SeqAIJ(Mat,Vec,PetscReal,MatSORType,PetscReal,PetscInt,PetscInt,Vec); 164 165 EXTERN PetscErrorCode MatSetColoring_SeqAIJ(Mat,ISColoring); 166 EXTERN PetscErrorCode MatSetValuesAdic_SeqAIJ(Mat,void*); 167 EXTERN PetscErrorCode MatSetValuesAdifor_SeqAIJ(Mat,PetscInt,void*); 168 169 EXTERN PetscErrorCode MatGetSymbolicTranspose_SeqAIJ(Mat,PetscInt *[],PetscInt *[]); 170 EXTERN PetscErrorCode MatGetSymbolicTransposeReduced_SeqAIJ(Mat,PetscInt,PetscInt,PetscInt *[],PetscInt *[]); 171 EXTERN PetscErrorCode MatRestoreSymbolicTranspose_SeqAIJ(Mat,PetscInt *[],PetscInt *[]); 172 EXTERN PetscErrorCode MatToSymmetricIJ_SeqAIJ(PetscInt,PetscInt*,PetscInt*,PetscInt,PetscInt,PetscInt**,PetscInt**); 173 EXTERN PetscErrorCode MatLUFactorSymbolic_SeqAIJ_inplace(Mat,Mat,IS,IS,const MatFactorInfo*); 174 EXTERN PetscErrorCode MatLUFactorSymbolic_SeqAIJ(Mat,Mat,IS,IS,const MatFactorInfo*); 175 EXTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_inplace(Mat,Mat,const MatFactorInfo*); 176 EXTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ(Mat,Mat,const MatFactorInfo*); 177 EXTERN PetscErrorCode MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(Mat,Mat,const MatFactorInfo*); 178 EXTERN PetscErrorCode MatLUFactor_SeqAIJ(Mat,IS,IS,const MatFactorInfo*); 179 EXTERN PetscErrorCode MatSolve_SeqAIJ_inplace(Mat,Vec,Vec); 180 EXTERN PetscErrorCode MatSolve_SeqAIJ(Mat,Vec,Vec); 181 EXTERN PetscErrorCode MatSolve_SeqAIJ_Inode_inplace(Mat,Vec,Vec); 182 EXTERN PetscErrorCode MatSolve_SeqAIJ_Inode(Mat,Vec,Vec); 183 EXTERN PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering_inplace(Mat,Vec,Vec); 184 EXTERN PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering(Mat,Vec,Vec); 185 EXTERN PetscErrorCode MatSolve_SeqAIJ_InplaceWithPerm(Mat,Vec,Vec); 186 EXTERN PetscErrorCode MatSolveAdd_SeqAIJ_inplace(Mat,Vec,Vec,Vec); 187 EXTERN PetscErrorCode MatSolveAdd_SeqAIJ(Mat,Vec,Vec,Vec); 188 EXTERN PetscErrorCode MatSolveTranspose_SeqAIJ_inplace(Mat,Vec,Vec); 189 EXTERN PetscErrorCode MatSolveTranspose_SeqAIJ(Mat,Vec,Vec); 190 EXTERN PetscErrorCode MatSolveTransposeAdd_SeqAIJ_inplace(Mat,Vec,Vec,Vec); 191 EXTERN PetscErrorCode MatSolveTransposeAdd_SeqAIJ(Mat,Vec,Vec,Vec); 192 EXTERN PetscErrorCode MatMatSolve_SeqAIJ_inplace(Mat,Mat,Mat); 193 EXTERN PetscErrorCode MatMatSolve_SeqAIJ(Mat,Mat,Mat); 194 EXTERN PetscErrorCode MatEqual_SeqAIJ(Mat A,Mat B,PetscBool * flg); 195 EXTERN PetscErrorCode MatFDColoringCreate_SeqAIJ(Mat,ISColoring,MatFDColoring); 196 EXTERN PetscErrorCode MatLoad_SeqAIJ(Mat,PetscViewer); 197 EXTERN PetscErrorCode RegisterApplyPtAPRoutines_Private(Mat); 198 EXTERN PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ(Mat,Mat,PetscReal,Mat*); 199 EXTERN PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ(Mat,Mat,Mat); 200 EXTERN PetscErrorCode MatPtAPSymbolic_SeqAIJ(Mat,Mat,PetscReal,Mat*); 201 EXTERN PetscErrorCode MatPtAPNumeric_SeqAIJ(Mat,Mat,Mat); 202 EXTERN PetscErrorCode MatPtAPSymbolic_SeqAIJ_SeqAIJ(Mat,Mat,PetscReal,Mat*); 203 EXTERN PetscErrorCode MatPtAPNumeric_SeqAIJ_SeqAIJ(Mat,Mat,Mat); 204 EXTERN PetscErrorCode MatMatMultTranspose_SeqAIJ_SeqAIJ(Mat,Mat,MatReuse,PetscReal,Mat*); 205 EXTERN PetscErrorCode MatMatMultTransposeSymbolic_SeqAIJ_SeqAIJ(Mat,Mat,PetscReal,Mat*); 206 EXTERN PetscErrorCode MatMatMultTransposeNumeric_SeqAIJ_SeqAIJ(Mat,Mat,Mat); 207 EXTERN PetscErrorCode MatSetValues_SeqAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const PetscScalar[],InsertMode); 208 EXTERN PetscErrorCode MatGetRow_SeqAIJ(Mat,PetscInt,PetscInt*,PetscInt**,PetscScalar**); 209 EXTERN PetscErrorCode MatRestoreRow_SeqAIJ(Mat,PetscInt,PetscInt*,PetscInt**,PetscScalar**); 210 EXTERN PetscErrorCode MatAXPY_SeqAIJ(Mat,PetscScalar,Mat,MatStructure); 211 EXTERN PetscErrorCode MatGetRowIJ_SeqAIJ(Mat,PetscInt,PetscBool ,PetscBool ,PetscInt*,PetscInt *[],PetscInt *[],PetscBool *); 212 EXTERN PetscErrorCode MatRestoreRowIJ_SeqAIJ(Mat,PetscInt,PetscBool ,PetscBool ,PetscInt *,PetscInt *[],PetscInt *[],PetscBool *); 213 EXTERN PetscErrorCode MatGetColumnIJ_SeqAIJ(Mat,PetscInt,PetscBool ,PetscBool ,PetscInt*,PetscInt *[],PetscInt *[],PetscBool *); 214 EXTERN PetscErrorCode MatRestoreColumnIJ_SeqAIJ(Mat,PetscInt,PetscBool ,PetscBool ,PetscInt *,PetscInt *[],PetscInt *[],PetscBool *); 215 EXTERN PetscErrorCode MatDestroy_SeqAIJ(Mat); 216 EXTERN PetscErrorCode MatView_SeqAIJ(Mat,PetscViewer); 217 218 EXTERN PetscErrorCode Mat_CheckInode(Mat,PetscBool ); 219 EXTERN PetscErrorCode Mat_CheckInode_FactorLU(Mat,PetscBool ); 220 221 EXTERN PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat,Mat,PetscInt*); 222 223 EXTERN_C_BEGIN 224 EXTERN PetscErrorCode PETSCMAT_DLLEXPORT MatConvert_SeqAIJ_SeqSBAIJ(Mat,const MatType,MatReuse,Mat*); 225 EXTERN PetscErrorCode PETSCMAT_DLLEXPORT MatConvert_SeqAIJ_SeqBAIJ(Mat,const MatType,MatReuse,Mat*); 226 EXTERN PetscErrorCode PETSCMAT_DLLEXPORT MatConvert_SeqAIJ_SeqAIJPERM(Mat,const MatType,MatReuse,Mat*); 227 EXTERN PetscErrorCode PETSCMAT_DLLEXPORT MatReorderForNonzeroDiagonal_SeqAIJ(Mat,PetscReal,IS,IS); 228 EXTERN PetscErrorCode PETSCMAT_DLLEXPORT MatMatMult_SeqAIJ_SeqAIJ(Mat,Mat,MatReuse,PetscReal,Mat*); 229 EXTERN PetscErrorCode PETSCMAT_DLLEXPORT MatCreate_SeqAIJ(Mat); 230 EXTERN_C_END 231 EXTERN PetscErrorCode PETSCMAT_DLLEXPORT MatAssemblyEnd_SeqAIJ(Mat,MatAssemblyType); 232 EXTERN PetscErrorCode PETSCMAT_DLLEXPORT MatDestroy_SeqAIJ(Mat); 233 234 235 /* 236 PetscSparseDenseMinusDot - The inner kernel of triangular solves and Gauss-Siedel smoothing. \sum_i xv[i] * r[xi[i]] for CSR storage 237 238 Input Parameters: 239 + nnz - the number of entries 240 . r - the array of vector values 241 . xv - the matrix values for the row 242 - xi - the column indices of the nonzeros in the row 243 244 Output Parameter: 245 . sum - negative the sum of results 246 247 PETSc compile flags: 248 + PETSC_KERNEL_USE_UNROLL_4 - don't use this; it changes nnz and hence is WRONG 249 - PETSC_KERNEL_USE_UNROLL_2 - 250 251 .seealso: PetscSparseDensePlusDot() 252 253 */ 254 #ifdef PETSC_KERNEL_USE_UNROLL_4 255 #define PetscSparseDenseMinusDot(sum,r,xv,xi,nnz) {\ 256 if (nnz > 0) {\ 257 switch (nnz & 0x3) {\ 258 case 3: sum -= *xv++ * r[*xi++];\ 259 case 2: sum -= *xv++ * r[*xi++];\ 260 case 1: sum -= *xv++ * r[*xi++];\ 261 nnz -= 4;}\ 262 while (nnz > 0) {\ 263 sum -= xv[0] * r[xi[0]] - xv[1] * r[xi[1]] -\ 264 xv[2] * r[xi[2]] - xv[3] * r[xi[3]];\ 265 xv += 4; xi += 4; nnz -= 4; }}} 266 267 #elif defined(PETSC_KERNEL_USE_UNROLL_2) 268 #define PetscSparseDenseMinusDot(sum,r,xv,xi,nnz) {\ 269 PetscInt __i,__i1,__i2;\ 270 for(__i=0;__i<nnz-1;__i+=2) {__i1 = xi[__i]; __i2=xi[__i+1];\ 271 sum -= (xv[__i]*r[__i1] + xv[__i+1]*r[__i2]);}\ 272 if (nnz & 0x1) sum -= xv[__i] * r[xi[__i]];} 273 274 #else 275 #define PetscSparseDenseMinusDot(sum,r,xv,xi,nnz) {\ 276 PetscInt __i;\ 277 for(__i=0;__i<nnz;__i++) sum -= xv[__i] * r[xi[__i]];} 278 #endif 279 280 281 282 /* 283 PetscSparseDensePlusDot - The inner kernel of matrix-vector product \sum_i xv[i] * r[xi[i]] for CSR storage 284 285 Input Parameters: 286 + nnz - the number of entries 287 . r - the array of vector values 288 . xv - the matrix values for the row 289 - xi - the column indices of the nonzeros in the row 290 291 Output Parameter: 292 . sum - the sum of results 293 294 PETSc compile flags: 295 + PETSC_KERNEL_USE_UNROLL_4 - don't use this; it changes nnz and hence is WRONG 296 - PETSC_KERNEL_USE_UNROLL_2 - 297 298 .seealso: PetscSparseDenseMinusDot() 299 300 */ 301 #ifdef PETSC_KERNEL_USE_UNROLL_4 302 #define PetscSparseDensePlusDot(sum,r,xv,xi,nnz) {\ 303 if (nnz > 0) {\ 304 switch (nnz & 0x3) {\ 305 case 3: sum += *xv++ * r[*xi++];\ 306 case 2: sum += *xv++ * r[*xi++];\ 307 case 1: sum += *xv++ * r[*xi++];\ 308 nnz -= 4;}\ 309 while (nnz > 0) {\ 310 sum += xv[0] * r[xi[0]] + xv[1] * r[xi[1]] +\ 311 xv[2] * r[xi[2]] + xv[3] * r[xi[3]];\ 312 xv += 4; xi += 4; nnz -= 4; }}} 313 314 #elif defined(PETSC_KERNEL_USE_UNROLL_2) 315 #define PetscSparseDensePlusDot(sum,r,xv,xi,nnz) {\ 316 PetscInt __i,__i1,__i2;\ 317 for(__i=0;__i<nnz-1;__i+=2) {__i1 = xi[__i]; __i2=xi[__i+1];\ 318 sum += (xv[__i]*r[__i1] + xv[__i+1]*r[__i2]);}\ 319 if (nnz & 0x1) sum += xv[__i] * r[xi[__i]];} 320 321 #else 322 #define PetscSparseDensePlusDot(sum,r,xv,xi,nnz) {\ 323 PetscInt __i;\ 324 for(__i=0;__i<nnz;__i++) sum += xv[__i] * r[xi[__i]];} 325 #endif 326 327 #endif 328