1 2 #ifndef __SELL_H 3 #define __SELL_H 4 5 #include <petsc/private/matimpl.h> 6 #include <petsc/private/hashmapi.h> 7 8 /* 9 For NVIDIA GPUs each slice should be padded to the boundary of 16 elements for best performance. 10 The optimal memory alignment in device memory is 128 bytes, 64 bytes, 32 bytes for double precision, single precision and half precision. 11 */ 12 #if defined(PETSC_HAVE_DEVICE) 13 #define DEVICE_MEM_ALIGN 16 14 #endif 15 16 /* 17 Struct header for SeqSELL matrix format 18 */ 19 #define SEQSELLHEADER(datatype) \ 20 PetscBool roworiented; /* if true, row-oriented input, default */ \ 21 PetscInt nonew; /* 1 don't add new nonzeros, -1 generate error on new */ \ 22 PetscInt nounused; /* -1 generate error on unused space */ \ 23 PetscBool singlemalloc; /* if true a, i, and j have been obtained with one big malloc */ \ 24 PetscInt maxallocmat; /* max allocated space for the matrix */ \ 25 PetscInt maxallocrow; /* max allocated space for each row */ \ 26 PetscInt nz; /* actual nonzeros */ \ 27 PetscInt rlenmax; /* max actual row length, rmax cannot exceed maxallocrow */ \ 28 PetscInt *rlen; /* actual length of each row (padding zeros excluded) */ \ 29 PetscBool free_rlen; /* free rlen array ? */ \ 30 PetscInt reallocs; /* number of mallocs done during MatSetValues() \ 31 as more values are set than were prealloced */ \ 32 PetscBool keepnonzeropattern; /* keeps matrix structure same in calls to MatZeroRows()*/ \ 33 PetscBool ignorezeroentries; \ 34 PetscBool free_colidx; /* free the column indices colidx when the matrix is destroyed */ \ 35 PetscBool free_val; /* free the numerical values when matrix is destroy */ \ 36 PetscInt *colidx; /* column index */ \ 37 PetscInt *diag; /* pointers to diagonal elements */ \ 38 PetscInt nonzerorowcnt; /* how many rows have nonzero entries */ \ 39 PetscBool free_diag; /* free diag ? */ \ 40 datatype *val; /* elements including nonzeros and padding zeros */ \ 41 PetscScalar *solve_work; /* work space used in MatSolve */ \ 42 IS row, col, icol; /* index sets, used for reorderings */ \ 43 PetscBool pivotinblocks; /* pivot inside factorization of each diagonal block */ \ 44 Mat parent; /* set if this matrix was formed with MatDuplicate(...,MAT_SHARE_NONZERO_PATTERN,....); 45 means that this shares some data structures with the parent including diag, ilen, imax, i, j */ \ 46 PetscInt *sliidx; /* slice index */ \ 47 PetscInt totalslices; /* total number of slices */ \ 48 PetscInt sliceheight; /* slice height */ \ 49 PetscReal fillratio; /* ratio of number of padded zeros over total number of elements */ \ 50 PetscReal avgslicewidth; /* average slice width */ \ 51 PetscInt maxslicewidth; /* maximum slice width */ \ 52 PetscInt *sliperm; /* slice permutation array, CUDA only */ \ 53 PetscInt totalblocks; /* total number of blocks, CUDA only */ \ 54 PetscInt *blockidx; /* block index, CUDA only */ \ 55 PetscInt *block_row_map; /* starting row of the current block, CUDA only */ \ 56 PetscInt chunksize; /* chunk size, CUDA only */ \ 57 PetscInt totalchunks; /* total number of chunks, CUDA only */ \ 58 PetscInt *chunk_slice_map; /* starting slice of the currect chunk, CUDA only */ \ 59 PetscInt *getrowcols; /* workarray for MatGetRow_SeqSELL */ \ 60 PetscScalar *getrowvals /* workarray for MatGetRow_SeqSELL */ 61 62 typedef struct { 63 SEQSELLHEADER(MatScalar); 64 MatScalar *saved_values; /* location for stashing nonzero values of matrix */ 65 PetscScalar *idiag, *mdiag, *ssor_work; /* inverse of diagonal entries, diagonal values and workspace for Eisenstat trick */ 66 PetscBool idiagvalid; /* current idiag[] and mdiag[] are valid */ 67 PetscScalar fshift, omega; /* last used omega and fshift */ 68 ISColoring coloring; /* set with MatADSetColoring() used by MatADSetValues() */ 69 } Mat_SeqSELL; 70 71 /* 72 Frees the arrays from the XSELLPACK matrix type 73 */ 74 static inline PetscErrorCode MatSeqXSELLFreeSELL(Mat AA, MatScalar **val, PetscInt **colidx) 75 { 76 Mat_SeqSELL *A = (Mat_SeqSELL *)AA->data; 77 if (A->singlemalloc) { 78 PetscCall(PetscFree2(*val, *colidx)); 79 } else { 80 if (A->free_val) PetscCall(PetscFree(*val)); 81 if (A->free_colidx) PetscCall(PetscFree(*colidx)); 82 } 83 return PETSC_SUCCESS; 84 } 85 86 #define MatSeqXSELLReallocateSELL(Amat, AM, BS2, WIDTH, SIDX, SH, SID, ROW, COL, COLIDX, VAL, CP, VP, NONEW, datatype, MUL) \ 87 if (WIDTH >= (SIDX[SID + 1] - SIDX[SID]) / SH) { \ 88 Mat_SeqSELL *Ain = (Mat_SeqSELL *)Amat->data; \ 89 /* there is no extra room in row, therefore enlarge 1 slice column */ \ 90 PetscInt new_size = Ain->maxallocmat + SH * MUL, *new_colidx; \ 91 datatype *new_val; \ 92 \ 93 PetscCheck(NONEW != -2, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "New nonzero at (%" PetscInt_FMT ",%" PetscInt_FMT ") caused a malloc\nUse MatSetOption(A, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE) to turn off this check", ROW, COL); \ 94 /* malloc new storage space */ \ 95 PetscCall(PetscMalloc2(BS2 *new_size, &new_val, BS2 *new_size, &new_colidx)); \ 96 \ 97 /* copy over old data into new slots by two steps: one step for data before the current slice and the other for the rest */ \ 98 PetscCall(PetscArraycpy(new_val, VAL, SIDX[SID + 1])); \ 99 PetscCall(PetscArraycpy(new_colidx, COLIDX, SIDX[SID + 1])); \ 100 PetscCall(PetscArraycpy(new_val + SIDX[SID + 1] + SH * MUL, VAL + SIDX[SID + 1], SIDX[Ain->totalslices] - SIDX[SID + 1])); \ 101 PetscCall(PetscArraycpy(new_colidx + SIDX[SID + 1] + SH * MUL, COLIDX + SIDX[SID + 1], SIDX[Ain->totalslices] - SIDX[SID + 1])); \ 102 /* update slice_idx */ \ 103 for (ii = SID + 1; ii <= Ain->totalslices; ii++) { SIDX[ii] += SH * MUL; } \ 104 /* update pointers. Notice that they point to the FIRST postion of the row */ \ 105 CP = new_colidx + SIDX[SID] + (ROW % SH); \ 106 VP = new_val + SIDX[SID] + (ROW % SH); \ 107 /* free up old matrix storage */ \ 108 PetscCall(MatSeqXSELLFreeSELL(A, &Ain->val, &Ain->colidx)); \ 109 Ain->val = (MatScalar *)new_val; \ 110 Ain->colidx = new_colidx; \ 111 Ain->singlemalloc = PETSC_TRUE; \ 112 Ain->maxallocmat = new_size; \ 113 Ain->reallocs++; \ 114 if (WIDTH >= Ain->maxallocrow) Ain->maxallocrow += MUL; \ 115 if (WIDTH >= Ain->rlenmax) Ain->rlenmax++; \ 116 } 117 118 #define MatSetValue_SeqSELL_Private(A, row, col, value, addv, orow, ocol, cp, vp, lastcol, low, high) \ 119 { \ 120 Mat_SeqSELL *a = (Mat_SeqSELL *)A->data; \ 121 found = PETSC_FALSE; \ 122 if (col <= lastcol) low = 0; \ 123 else high = a->rlen[row]; \ 124 lastcol = col; \ 125 while (high - low > 5) { \ 126 t = (low + high) / 2; \ 127 if (*(cp + a->sliceheight * t) > col) high = t; \ 128 else low = t; \ 129 } \ 130 for (_i = low; _i < high; _i++) { \ 131 if (*(cp + a->sliceheight * _i) > col) break; \ 132 if (*(cp + a->sliceheight * _i) == col) { \ 133 if (addv == ADD_VALUES) *(vp + a->sliceheight * _i) += value; \ 134 else *(vp + a->sliceheight * _i) = value; \ 135 found = PETSC_TRUE; \ 136 break; \ 137 } \ 138 } \ 139 if (!found) { \ 140 PetscCheck(a->nonew != -1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero at global row/column (%" PetscInt_FMT ", %" PetscInt_FMT ") into matrix", orow, ocol); \ 141 if (a->nonew != 1 && !(value == 0.0 && a->ignorezeroentries) && a->rlen[row] >= (a->sliidx[row / a->sliceheight + 1] - a->sliidx[row / a->sliceheight]) / a->sliceheight) { \ 142 /* there is no extra room in row, therefore enlarge 1 slice column */ \ 143 if (a->maxallocmat < a->sliidx[a->totalslices] + a->sliceheight) { \ 144 /* allocates a larger array for the XSELL matrix types; only extend the current slice by one more column. */ \ 145 PetscInt new_size = a->maxallocmat + a->sliceheight, *new_colidx; \ 146 MatScalar *new_val; \ 147 PetscCheck(a->nonew != -2, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "New nonzero at (%" PetscInt_FMT ",%" PetscInt_FMT ") caused a malloc\nUse MatSetOption(A, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE) to turn off this check", orow, ocol); \ 148 /* malloc new storage space */ \ 149 PetscCall(PetscMalloc2(new_size, &new_val, new_size, &new_colidx)); \ 150 /* copy over old data into new slots by two steps: one step for data before the current slice and the other for the rest */ \ 151 PetscCall(PetscArraycpy(new_val, a->val, a->sliidx[row / a->sliceheight + 1])); \ 152 PetscCall(PetscArraycpy(new_colidx, a->colidx, a->sliidx[row / a->sliceheight + 1])); \ 153 PetscCall(PetscArraycpy(new_val + a->sliidx[row / a->sliceheight + 1] + a->sliceheight, a->val + a->sliidx[row / a->sliceheight + 1], a->sliidx[a->totalslices] - a->sliidx[row / a->sliceheight + 1])); \ 154 PetscCall(PetscArraycpy(new_colidx + a->sliidx[row / a->sliceheight + 1] + a->sliceheight, a->colidx + a->sliidx[row / a->sliceheight + 1], a->sliidx[a->totalslices] - a->sliidx[row / a->sliceheight + 1])); \ 155 /* update pointers. Notice that they point to the FIRST postion of the row */ \ 156 cp = new_colidx + a->sliidx[row / a->sliceheight] + (row % a->sliceheight); \ 157 vp = new_val + a->sliidx[row / a->sliceheight] + (row % a->sliceheight); \ 158 /* free up old matrix storage */ \ 159 PetscCall(MatSeqXSELLFreeSELL(A, &a->val, &a->colidx)); \ 160 a->val = (MatScalar *)new_val; \ 161 a->colidx = new_colidx; \ 162 a->singlemalloc = PETSC_TRUE; \ 163 a->maxallocmat = new_size; \ 164 a->reallocs++; \ 165 } else { \ 166 /* no need to reallocate, just shift the following slices to create space for the added slice column */ \ 167 PetscCall(PetscArraymove(a->val + a->sliidx[row / a->sliceheight + 1] + a->sliceheight, a->val + a->sliidx[row / a->sliceheight + 1], a->sliidx[a->totalslices] - a->sliidx[row / a->sliceheight + 1])); \ 168 PetscCall(PetscArraymove(a->colidx + a->sliidx[row / a->sliceheight + 1] + a->sliceheight, a->colidx + a->sliidx[row / a->sliceheight + 1], a->sliidx[a->totalslices] - a->sliidx[row / a->sliceheight + 1])); \ 169 } \ 170 /* update slice_idx */ \ 171 for (ii = row / a->sliceheight + 1; ii <= a->totalslices; ii++) a->sliidx[ii] += a->sliceheight; \ 172 if (a->rlen[row] >= a->maxallocrow) a->maxallocrow++; \ 173 if (a->rlen[row] >= a->rlenmax) a->rlenmax++; \ 174 } \ 175 /* shift up all the later entries in this row */ \ 176 for (ii = a->rlen[row] - 1; ii >= _i; ii--) { \ 177 *(cp + a->sliceheight * (ii + 1)) = *(cp + a->sliceheight * ii); \ 178 *(vp + a->sliceheight * (ii + 1)) = *(vp + a->sliceheight * ii); \ 179 } \ 180 *(cp + a->sliceheight * _i) = col; \ 181 *(vp + a->sliceheight * _i) = value; \ 182 a->nz++; \ 183 a->rlen[row]++; \ 184 A->nonzerostate++; \ 185 low = _i + 1; \ 186 high++; \ 187 } \ 188 } 189 190 PETSC_INTERN PetscErrorCode MatSeqSELLSetPreallocation_SeqSELL(Mat, PetscInt, const PetscInt[]); 191 PETSC_INTERN PetscErrorCode MatMult_SeqSELL(Mat, Vec, Vec); 192 PETSC_INTERN PetscErrorCode MatMultAdd_SeqSELL(Mat, Vec, Vec, Vec); 193 PETSC_INTERN PetscErrorCode MatMultTranspose_SeqSELL(Mat, Vec, Vec); 194 PETSC_INTERN PetscErrorCode MatMultTransposeAdd_SeqSELL(Mat, Vec, Vec, Vec); 195 PETSC_INTERN PetscErrorCode MatMissingDiagonal_SeqSELL(Mat, PetscBool *, PetscInt *); 196 PETSC_INTERN PetscErrorCode MatMarkDiagonal_SeqSELL(Mat); 197 PETSC_INTERN PetscErrorCode MatInvertDiagonal_SeqSELL(Mat, PetscScalar, PetscScalar); 198 PETSC_INTERN PetscErrorCode MatZeroEntries_SeqSELL(Mat); 199 PETSC_INTERN PetscErrorCode MatDestroy_SeqSELL(Mat); 200 PETSC_INTERN PetscErrorCode MatSetOption_SeqSELL(Mat, MatOption, PetscBool); 201 PETSC_INTERN PetscErrorCode MatGetDiagonal_SeqSELL(Mat, Vec v); 202 PETSC_INTERN PetscErrorCode MatGetValues_SeqSELL(Mat, PetscInt, const PetscInt[], PetscInt, const PetscInt[], PetscScalar[]); 203 PETSC_INTERN PetscErrorCode MatView_SeqSELL(Mat, PetscViewer); 204 PETSC_INTERN PetscErrorCode MatAssemblyEnd_SeqSELL(Mat, MatAssemblyType); 205 PETSC_INTERN PetscErrorCode MatGetInfo_SeqSELL(Mat, MatInfoType, MatInfo *); 206 PETSC_INTERN PetscErrorCode MatSetValues_SeqSELL(Mat, PetscInt, const PetscInt[], PetscInt, const PetscInt[], const PetscScalar[], InsertMode); 207 PETSC_INTERN PetscErrorCode MatCopy_SeqSELL(Mat, Mat, MatStructure); 208 PETSC_INTERN PetscErrorCode MatSetUp_SeqSELL(Mat); 209 PETSC_INTERN PetscErrorCode MatSeqSELLGetArray_SeqSELL(Mat, PetscScalar *[]); 210 PETSC_INTERN PetscErrorCode MatSeqSELLRestoreArray_SeqSELL(Mat, PetscScalar *[]); 211 PETSC_INTERN PetscErrorCode MatShift_SeqSELL(Mat, PetscScalar); 212 PETSC_INTERN PetscErrorCode MatSOR_SeqSELL(Mat, Vec, PetscReal, MatSORType, PetscReal, PetscInt, PetscInt, Vec); 213 PETSC_EXTERN PetscErrorCode MatCreate_SeqSELL(Mat); 214 PETSC_INTERN PetscErrorCode MatDuplicate_SeqSELL(Mat, MatDuplicateOption, Mat *); 215 PETSC_INTERN PetscErrorCode MatEqual_SeqSELL(Mat, Mat, PetscBool *); 216 PETSC_INTERN PetscErrorCode MatSeqSELLInvalidateDiagonal(Mat); 217 PETSC_INTERN PetscErrorCode MatConvert_SeqSELL_SeqAIJ(Mat, MatType, MatReuse, Mat *); 218 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqSELL(Mat, MatType, MatReuse, Mat *); 219 PETSC_INTERN PetscErrorCode MatFDColoringCreate_SeqSELL(Mat, ISColoring, MatFDColoring); 220 PETSC_INTERN PetscErrorCode MatFDColoringSetUp_SeqSELL(Mat, ISColoring, MatFDColoring); 221 PETSC_INTERN PetscErrorCode MatGetColumnIJ_SeqSELL_Color(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscInt *[], PetscBool *); 222 PETSC_INTERN PetscErrorCode MatRestoreColumnIJ_SeqSELL_Color(Mat, PetscInt, PetscBool, PetscBool, PetscInt *, const PetscInt *[], const PetscInt *[], PetscInt *[], PetscBool *); 223 PETSC_INTERN PetscErrorCode MatConjugate_SeqSELL(Mat A); 224 PETSC_INTERN PetscErrorCode MatScale_SeqSELL(Mat, PetscScalar); 225 PETSC_INTERN PetscErrorCode MatDiagonalScale_SeqSELL(Mat, Vec, Vec); 226 #endif 227