Actual source code: mkl_pardiso.c
1: #include <../src/mat/impls/aij/seq/aij.h>
2: #include <../src/mat/impls/sbaij/seq/sbaij.h>
3: #include <../src/mat/impls/dense/seq/dense.h>
5: #if defined(PETSC_HAVE_MKL_INTEL_ILP64)
6: #define MKL_ILP64
7: #endif
8: #include <mkl_pardiso.h>
10: PETSC_EXTERN void PetscSetMKL_PARDISOThreads(int);
12: /*
13: * Possible mkl_pardiso phases that controls the execution of the solver.
14: * For more information check mkl_pardiso manual.
15: */
16: #define JOB_ANALYSIS 11
17: #define JOB_ANALYSIS_NUMERICAL_FACTORIZATION 12
18: #define JOB_ANALYSIS_NUMERICAL_FACTORIZATION_SOLVE_ITERATIVE_REFINEMENT 13
19: #define JOB_NUMERICAL_FACTORIZATION 22
20: #define JOB_NUMERICAL_FACTORIZATION_SOLVE_ITERATIVE_REFINEMENT 23
21: #define JOB_SOLVE_ITERATIVE_REFINEMENT 33
22: #define JOB_SOLVE_FORWARD_SUBSTITUTION 331
23: #define JOB_SOLVE_DIAGONAL_SUBSTITUTION 332
24: #define JOB_SOLVE_BACKWARD_SUBSTITUTION 333
25: #define JOB_RELEASE_OF_LU_MEMORY 0
26: #define JOB_RELEASE_OF_ALL_MEMORY -1
28: #define IPARM_SIZE 64
30: #if defined(PETSC_USE_64BIT_INDICES)
31: #if defined(PETSC_HAVE_MKL_INTEL_ILP64)
32: #define INT_TYPE long long int
33: #define MKL_PARDISO pardiso
34: #define MKL_PARDISO_INIT pardisoinit
35: #else
36: /* this is the case where the MKL BLAS/LAPACK are 32-bit integers but the 64-bit integer version of
37: of PARDISO code is used; hence the need for the 64 below*/
38: #define INT_TYPE long long int
39: #define MKL_PARDISO pardiso_64
40: #define MKL_PARDISO_INIT pardiso_64init
41: void pardiso_64init(void *pt, INT_TYPE *mtype, INT_TYPE iparm[])
42: {
43: PetscBLASInt iparm_copy[IPARM_SIZE], mtype_copy;
45: PetscCallVoid(PetscBLASIntCast(*mtype, &mtype_copy));
46: pardisoinit(pt, &mtype_copy, iparm_copy);
47: for (PetscInt i = 0; i < IPARM_SIZE; i++) iparm[i] = iparm_copy[i];
48: }
49: #endif
50: #else
51: #define INT_TYPE int
52: #define MKL_PARDISO pardiso
53: #define MKL_PARDISO_INIT pardisoinit
54: #endif
56: #define PetscCallPardiso(f) PetscStackCallExternalVoid("MKL_PARDISO", f);
58: /*
59: Internal data structure.
60: */
61: typedef struct {
62: /* Configuration vector*/
63: INT_TYPE iparm[IPARM_SIZE];
65: /*
66: Internal MKL PARDISO memory location.
67: After the first call to MKL PARDISO do not modify pt, as that could cause a serious memory leak.
68: */
69: void *pt[IPARM_SIZE];
71: /* Basic MKL PARDISO info */
72: INT_TYPE phase, maxfct, mnum, mtype, n, nrhs, msglvl, err;
74: /* Matrix structure*/
75: void *a;
76: INT_TYPE *ia, *ja;
78: /* Number of non-zero elements*/
79: INT_TYPE nz;
81: /* Row permutaton vector*/
82: INT_TYPE *perm;
84: /* Define if matrix preserves sparse structure.*/
85: MatStructure matstruc;
87: PetscBool needsym;
88: PetscBool freeaij;
90: /* Schur complement */
91: PetscScalar *schur;
92: PetscInt schur_size;
93: PetscInt *schur_idxs;
94: PetscScalar *schur_work;
95: PetscBLASInt schur_work_size;
96: PetscBool solve_interior;
98: /* True if MKL PARDISO function have been used. */
99: PetscBool CleanUp;
101: /* Conversion to a format suitable for MKL */
102: PetscErrorCode (*Convert)(Mat, PetscBool, MatReuse, PetscBool *, INT_TYPE *, INT_TYPE **, INT_TYPE **, PetscScalar **);
103: } Mat_MKL_PARDISO;
105: static PetscErrorCode MatMKLPardiso_Convert_seqsbaij(Mat A, PetscBool sym, MatReuse reuse, PetscBool *free, INT_TYPE *nnz, INT_TYPE **r, INT_TYPE **c, PetscScalar **v)
106: {
107: Mat_SeqSBAIJ *aa = (Mat_SeqSBAIJ *)A->data;
108: PetscInt bs = A->rmap->bs, i;
110: PetscFunctionBegin;
111: PetscCheck(sym, PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "This should not happen");
112: *v = aa->a;
113: if (bs == 1) { /* already in the correct format */
114: /* though PetscInt and INT_TYPE are of the same size since they are defined differently the Intel compiler requires a cast */
115: *r = (INT_TYPE *)aa->i;
116: *c = (INT_TYPE *)aa->j;
117: *nnz = (INT_TYPE)aa->nz;
118: *free = PETSC_FALSE;
119: } else if (reuse == MAT_INITIAL_MATRIX) {
120: PetscInt m = A->rmap->n, nz = aa->nz;
121: PetscInt *row, *col;
122: PetscCall(PetscMalloc2(m + 1, &row, nz, &col));
123: for (i = 0; i < m + 1; i++) row[i] = aa->i[i] + 1;
124: for (i = 0; i < nz; i++) col[i] = aa->j[i] + 1;
125: *r = (INT_TYPE *)row;
126: *c = (INT_TYPE *)col;
127: *nnz = (INT_TYPE)nz;
128: *free = PETSC_TRUE;
129: }
130: PetscFunctionReturn(PETSC_SUCCESS);
131: }
133: static PetscErrorCode MatMKLPardiso_Convert_seqbaij(Mat A, PetscBool sym, MatReuse reuse, PetscBool *free, INT_TYPE *nnz, INT_TYPE **r, INT_TYPE **c, PetscScalar **v)
134: {
135: Mat_SeqBAIJ *aa = (Mat_SeqBAIJ *)A->data;
136: PetscInt bs = A->rmap->bs, i;
138: PetscFunctionBegin;
139: if (!sym) {
140: *v = aa->a;
141: if (bs == 1) { /* already in the correct format */
142: /* though PetscInt and INT_TYPE are of the same size since they are defined differently the Intel compiler requires a cast */
143: *r = (INT_TYPE *)aa->i;
144: *c = (INT_TYPE *)aa->j;
145: *nnz = (INT_TYPE)aa->nz;
146: *free = PETSC_FALSE;
147: PetscFunctionReturn(PETSC_SUCCESS);
148: } else if (reuse == MAT_INITIAL_MATRIX) {
149: PetscInt m = A->rmap->n, nz = aa->nz;
150: PetscInt *row, *col;
151: PetscCall(PetscMalloc2(m + 1, &row, nz, &col));
152: for (i = 0; i < m + 1; i++) row[i] = aa->i[i] + 1;
153: for (i = 0; i < nz; i++) col[i] = aa->j[i] + 1;
154: *r = (INT_TYPE *)row;
155: *c = (INT_TYPE *)col;
156: *nnz = (INT_TYPE)nz;
157: }
158: *free = PETSC_TRUE;
159: } else {
160: SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_PLIB, "This should not happen");
161: }
162: PetscFunctionReturn(PETSC_SUCCESS);
163: }
165: static PetscErrorCode MatMKLPardiso_Convert_seqaij(Mat A, PetscBool sym, MatReuse reuse, PetscBool *free, INT_TYPE *nnz, INT_TYPE **r, INT_TYPE **c, PetscScalar **v)
166: {
167: Mat_SeqAIJ *aa = (Mat_SeqAIJ *)A->data;
168: PetscScalar *aav;
170: PetscFunctionBegin;
171: PetscCall(MatSeqAIJGetArrayRead(A, (const PetscScalar **)&aav));
172: if (!sym) { /* already in the correct format */
173: *v = aav;
174: *r = (INT_TYPE *)aa->i;
175: *c = (INT_TYPE *)aa->j;
176: *nnz = (INT_TYPE)aa->nz;
177: *free = PETSC_FALSE;
178: } else if (reuse == MAT_INITIAL_MATRIX) { /* need to get the triangular part */
179: PetscScalar *vals, *vv;
180: PetscInt *row, *col, *jj;
181: PetscInt m = A->rmap->n, nz, i;
183: nz = 0;
184: for (i = 0; i < m; i++) nz += aa->i[i + 1] - aa->diag[i];
185: PetscCall(PetscMalloc2(m + 1, &row, nz, &col));
186: PetscCall(PetscMalloc1(nz, &vals));
187: jj = col;
188: vv = vals;
190: row[0] = 0;
191: for (i = 0; i < m; i++) {
192: PetscInt *aj = aa->j + aa->diag[i];
193: PetscScalar *av = aav + aa->diag[i];
194: PetscInt rl = aa->i[i + 1] - aa->diag[i], j;
196: for (j = 0; j < rl; j++) {
197: *jj = *aj;
198: jj++;
199: aj++;
200: *vv = *av;
201: vv++;
202: av++;
203: }
204: row[i + 1] = row[i] + rl;
205: }
206: *v = vals;
207: *r = (INT_TYPE *)row;
208: *c = (INT_TYPE *)col;
209: *nnz = (INT_TYPE)nz;
210: *free = PETSC_TRUE;
211: } else {
212: PetscScalar *vv;
213: PetscInt m = A->rmap->n, i;
215: vv = *v;
216: for (i = 0; i < m; i++) {
217: PetscScalar *av = aav + aa->diag[i];
218: PetscInt rl = aa->i[i + 1] - aa->diag[i], j;
219: for (j = 0; j < rl; j++) {
220: *vv = *av;
221: vv++;
222: av++;
223: }
224: }
225: *free = PETSC_TRUE;
226: }
227: PetscCall(MatSeqAIJRestoreArrayRead(A, (const PetscScalar **)&aav));
228: PetscFunctionReturn(PETSC_SUCCESS);
229: }
231: static PetscErrorCode MatMKLPardisoSolveSchur_Private(Mat F, PetscScalar *B, PetscScalar *X)
232: {
233: Mat_MKL_PARDISO *mpardiso = (Mat_MKL_PARDISO *)F->data;
234: Mat S, Xmat, Bmat;
235: MatFactorSchurStatus schurstatus;
237: PetscFunctionBegin;
238: PetscCall(MatFactorGetSchurComplement(F, &S, &schurstatus));
239: PetscCheck(X != B || schurstatus != MAT_FACTOR_SCHUR_INVERTED, PETSC_COMM_SELF, PETSC_ERR_SUP, "X and B cannot point to the same address");
240: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, mpardiso->schur_size, mpardiso->nrhs, B, &Bmat));
241: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, mpardiso->schur_size, mpardiso->nrhs, X, &Xmat));
242: PetscCall(MatSetType(Bmat, ((PetscObject)S)->type_name));
243: PetscCall(MatSetType(Xmat, ((PetscObject)S)->type_name));
244: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
245: PetscCall(MatBindToCPU(Xmat, S->boundtocpu));
246: PetscCall(MatBindToCPU(Bmat, S->boundtocpu));
247: #endif
249: #if defined(PETSC_USE_COMPLEX)
250: PetscCheck(mpardiso->iparm[12 - 1] != 1, PetscObjectComm((PetscObject)F), PETSC_ERR_SUP, "Hermitian solve not implemented yet");
251: #endif
253: switch (schurstatus) {
254: case MAT_FACTOR_SCHUR_FACTORED:
255: if (!mpardiso->iparm[12 - 1]) {
256: PetscCall(MatMatSolve(S, Bmat, Xmat));
257: } else { /* transpose solve */
258: PetscCall(MatMatSolveTranspose(S, Bmat, Xmat));
259: }
260: break;
261: case MAT_FACTOR_SCHUR_INVERTED:
262: PetscCall(MatProductCreateWithMat(S, Bmat, NULL, Xmat));
263: if (!mpardiso->iparm[12 - 1]) {
264: PetscCall(MatProductSetType(Xmat, MATPRODUCT_AB));
265: } else { /* transpose solve */
266: PetscCall(MatProductSetType(Xmat, MATPRODUCT_AtB));
267: }
268: PetscCall(MatProductSetFromOptions(Xmat));
269: PetscCall(MatProductSymbolic(Xmat));
270: PetscCall(MatProductNumeric(Xmat));
271: PetscCall(MatProductClear(Xmat));
272: break;
273: default:
274: SETERRQ(PetscObjectComm((PetscObject)F), PETSC_ERR_SUP, "Unhandled MatFactorSchurStatus %d", (int)F->schur_status);
275: break;
276: }
277: PetscCall(MatFactorRestoreSchurComplement(F, &S, schurstatus));
278: PetscCall(MatDestroy(&Bmat));
279: PetscCall(MatDestroy(&Xmat));
280: PetscFunctionReturn(PETSC_SUCCESS);
281: }
283: static PetscErrorCode MatFactorSetSchurIS_MKL_PARDISO(Mat F, IS is)
284: {
285: Mat_MKL_PARDISO *mpardiso = (Mat_MKL_PARDISO *)F->data;
286: const PetscScalar *arr;
287: const PetscInt *idxs;
288: PetscInt size, i;
289: PetscMPIInt csize;
290: PetscBool sorted;
292: PetscFunctionBegin;
293: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)F), &csize));
294: PetscCheck(csize <= 1, PETSC_COMM_SELF, PETSC_ERR_SUP, "MKL PARDISO parallel Schur complements not yet supported from PETSc");
295: PetscCall(ISSorted(is, &sorted));
296: PetscCheck(sorted, PETSC_COMM_SELF, PETSC_ERR_SUP, "IS for MKL PARDISO Schur complements needs to be sorted");
297: PetscCall(ISGetLocalSize(is, &size));
298: PetscCall(PetscFree(mpardiso->schur_work));
299: PetscCall(PetscBLASIntCast(PetscMax(mpardiso->n, 2 * size), &mpardiso->schur_work_size));
300: PetscCall(PetscMalloc1(mpardiso->schur_work_size, &mpardiso->schur_work));
301: PetscCall(MatDestroy(&F->schur));
302: PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, size, size, NULL, &F->schur));
303: PetscCall(MatDenseGetArrayRead(F->schur, &arr));
304: mpardiso->schur = (PetscScalar *)arr;
305: mpardiso->schur_size = size;
306: PetscCall(MatDenseRestoreArrayRead(F->schur, &arr));
307: if (mpardiso->mtype == 2) PetscCall(MatSetOption(F->schur, MAT_SPD, PETSC_TRUE));
309: PetscCall(PetscFree(mpardiso->schur_idxs));
310: PetscCall(PetscMalloc1(size, &mpardiso->schur_idxs));
311: PetscCall(PetscArrayzero(mpardiso->perm, mpardiso->n));
312: PetscCall(ISGetIndices(is, &idxs));
313: PetscCall(PetscArraycpy(mpardiso->schur_idxs, idxs, size));
314: for (i = 0; i < size; i++) mpardiso->perm[idxs[i]] = 1;
315: PetscCall(ISRestoreIndices(is, &idxs));
316: if (size) { /* turn on Schur switch if the set of indices is not empty */
317: mpardiso->iparm[36 - 1] = 2;
318: }
319: PetscFunctionReturn(PETSC_SUCCESS);
320: }
322: static PetscErrorCode MatDestroy_MKL_PARDISO(Mat A)
323: {
324: Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)A->data;
326: PetscFunctionBegin;
327: if (mat_mkl_pardiso->CleanUp) {
328: mat_mkl_pardiso->phase = JOB_RELEASE_OF_ALL_MEMORY;
330: PetscCallPardiso(MKL_PARDISO(mat_mkl_pardiso->pt, &mat_mkl_pardiso->maxfct, &mat_mkl_pardiso->mnum, &mat_mkl_pardiso->mtype, &mat_mkl_pardiso->phase, &mat_mkl_pardiso->n, NULL, NULL, NULL, NULL, &mat_mkl_pardiso->nrhs, mat_mkl_pardiso->iparm,
331: &mat_mkl_pardiso->msglvl, NULL, NULL, &mat_mkl_pardiso->err));
332: }
333: PetscCall(PetscFree(mat_mkl_pardiso->perm));
334: PetscCall(PetscFree(mat_mkl_pardiso->schur_work));
335: PetscCall(PetscFree(mat_mkl_pardiso->schur_idxs));
336: if (mat_mkl_pardiso->freeaij) {
337: PetscCall(PetscFree2(mat_mkl_pardiso->ia, mat_mkl_pardiso->ja));
338: if (mat_mkl_pardiso->iparm[34] == 1) PetscCall(PetscFree(mat_mkl_pardiso->a));
339: }
340: PetscCall(PetscFree(A->data));
342: /* clear composed functions */
343: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatFactorGetSolverType_C", NULL));
344: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatFactorSetSchurIS_C", NULL));
345: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatMkl_PardisoSetCntl_C", NULL));
346: PetscFunctionReturn(PETSC_SUCCESS);
347: }
349: static PetscErrorCode MatMKLPardisoScatterSchur_Private(Mat_MKL_PARDISO *mpardiso, PetscScalar *whole, PetscScalar *schur, PetscBool reduce)
350: {
351: PetscFunctionBegin;
352: if (reduce) { /* data given for the whole matrix */
353: PetscInt i, m = 0, p = 0;
354: for (i = 0; i < mpardiso->nrhs; i++) {
355: PetscInt j;
356: for (j = 0; j < mpardiso->schur_size; j++) schur[p + j] = whole[m + mpardiso->schur_idxs[j]];
357: m += mpardiso->n;
358: p += mpardiso->schur_size;
359: }
360: } else { /* from Schur to whole */
361: PetscInt i, m = 0, p = 0;
362: for (i = 0; i < mpardiso->nrhs; i++) {
363: PetscInt j;
364: for (j = 0; j < mpardiso->schur_size; j++) whole[m + mpardiso->schur_idxs[j]] = schur[p + j];
365: m += mpardiso->n;
366: p += mpardiso->schur_size;
367: }
368: }
369: PetscFunctionReturn(PETSC_SUCCESS);
370: }
372: static PetscErrorCode MatSolve_MKL_PARDISO(Mat A, Vec b, Vec x)
373: {
374: Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)A->data;
375: PetscScalar *xarray;
376: const PetscScalar *barray;
378: PetscFunctionBegin;
379: mat_mkl_pardiso->nrhs = 1;
380: PetscCall(VecGetArrayWrite(x, &xarray));
381: PetscCall(VecGetArrayRead(b, &barray));
383: if (!mat_mkl_pardiso->schur) mat_mkl_pardiso->phase = JOB_SOLVE_ITERATIVE_REFINEMENT;
384: else mat_mkl_pardiso->phase = JOB_SOLVE_FORWARD_SUBSTITUTION;
386: if (barray == xarray) { /* if the two vectors share the same memory */
387: PetscScalar *work;
388: if (!mat_mkl_pardiso->schur_work) {
389: PetscCall(PetscMalloc1(mat_mkl_pardiso->n, &work));
390: } else {
391: work = mat_mkl_pardiso->schur_work;
392: }
393: mat_mkl_pardiso->iparm[6 - 1] = 1;
394: PetscCallPardiso(MKL_PARDISO(mat_mkl_pardiso->pt, &mat_mkl_pardiso->maxfct, &mat_mkl_pardiso->mnum, &mat_mkl_pardiso->mtype, &mat_mkl_pardiso->phase, &mat_mkl_pardiso->n, mat_mkl_pardiso->a, mat_mkl_pardiso->ia, mat_mkl_pardiso->ja, NULL,
395: &mat_mkl_pardiso->nrhs, mat_mkl_pardiso->iparm, &mat_mkl_pardiso->msglvl, (void *)xarray, (void *)work, &mat_mkl_pardiso->err));
396: if (!mat_mkl_pardiso->schur_work) PetscCall(PetscFree(work));
397: } else {
398: mat_mkl_pardiso->iparm[6 - 1] = 0;
399: PetscCallPardiso(MKL_PARDISO(mat_mkl_pardiso->pt, &mat_mkl_pardiso->maxfct, &mat_mkl_pardiso->mnum, &mat_mkl_pardiso->mtype, &mat_mkl_pardiso->phase, &mat_mkl_pardiso->n, mat_mkl_pardiso->a, mat_mkl_pardiso->ia, mat_mkl_pardiso->ja,
400: mat_mkl_pardiso->perm, &mat_mkl_pardiso->nrhs, mat_mkl_pardiso->iparm, &mat_mkl_pardiso->msglvl, (void *)barray, (void *)xarray, &mat_mkl_pardiso->err));
401: }
402: PetscCall(VecRestoreArrayRead(b, &barray));
404: PetscCheck(mat_mkl_pardiso->err >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error reported by MKL PARDISO: err=%" PetscInt_FMT ". Please check manual", (PetscInt)mat_mkl_pardiso->err);
406: if (mat_mkl_pardiso->schur) { /* solve Schur complement and expand solution */
407: if (!mat_mkl_pardiso->solve_interior) {
408: PetscInt shift = mat_mkl_pardiso->schur_size;
410: PetscCall(MatFactorFactorizeSchurComplement(A));
411: /* if inverted, uses BLAS *MM subroutines, otherwise LAPACK *TRS */
412: if (A->schur_status != MAT_FACTOR_SCHUR_INVERTED) shift = 0;
414: /* solve Schur complement */
415: PetscCall(MatMKLPardisoScatterSchur_Private(mat_mkl_pardiso, xarray, mat_mkl_pardiso->schur_work, PETSC_TRUE));
416: PetscCall(MatMKLPardisoSolveSchur_Private(A, mat_mkl_pardiso->schur_work, mat_mkl_pardiso->schur_work + shift));
417: PetscCall(MatMKLPardisoScatterSchur_Private(mat_mkl_pardiso, xarray, mat_mkl_pardiso->schur_work + shift, PETSC_FALSE));
418: } else { /* if we are solving for the interior problem, any value in barray[schur] forward-substituted to xarray[schur] will be neglected */
419: PetscInt i;
420: for (i = 0; i < mat_mkl_pardiso->schur_size; i++) xarray[mat_mkl_pardiso->schur_idxs[i]] = 0.;
421: }
423: /* expansion phase */
424: mat_mkl_pardiso->iparm[6 - 1] = 1;
425: mat_mkl_pardiso->phase = JOB_SOLVE_BACKWARD_SUBSTITUTION;
426: PetscCallPardiso(MKL_PARDISO(mat_mkl_pardiso->pt, &mat_mkl_pardiso->maxfct, &mat_mkl_pardiso->mnum, &mat_mkl_pardiso->mtype, &mat_mkl_pardiso->phase, &mat_mkl_pardiso->n, mat_mkl_pardiso->a, mat_mkl_pardiso->ia, mat_mkl_pardiso->ja,
427: mat_mkl_pardiso->perm, &mat_mkl_pardiso->nrhs, mat_mkl_pardiso->iparm, &mat_mkl_pardiso->msglvl, (void *)xarray, (void *)mat_mkl_pardiso->schur_work, /* according to the specs, the solution vector is always used */
428: &mat_mkl_pardiso->err));
429: PetscCheck(mat_mkl_pardiso->err >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error reported by MKL PARDISO: err=%" PetscInt_FMT ". Please check manual", (PetscInt)mat_mkl_pardiso->err);
430: mat_mkl_pardiso->iparm[6 - 1] = 0;
431: }
432: PetscCall(VecRestoreArrayWrite(x, &xarray));
433: mat_mkl_pardiso->CleanUp = PETSC_TRUE;
434: PetscFunctionReturn(PETSC_SUCCESS);
435: }
437: static PetscErrorCode MatForwardSolve_MKL_PARDISO(Mat A, Vec b, Vec x)
438: {
439: Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)A->data;
440: PetscScalar *xarray;
441: const PetscScalar *barray;
443: PetscFunctionBegin;
444: PetscCheck(!mat_mkl_pardiso->schur, PETSC_COMM_SELF, PETSC_ERR_SUP, "Forward substitution not supported with Schur complement");
446: mat_mkl_pardiso->nrhs = 1;
447: PetscCall(VecGetArrayWrite(x, &xarray));
448: PetscCall(VecGetArrayRead(b, &barray));
450: mat_mkl_pardiso->phase = JOB_SOLVE_FORWARD_SUBSTITUTION;
452: PetscCallPardiso(MKL_PARDISO(mat_mkl_pardiso->pt, &mat_mkl_pardiso->maxfct, &mat_mkl_pardiso->mnum, &mat_mkl_pardiso->mtype, &mat_mkl_pardiso->phase, &mat_mkl_pardiso->n, mat_mkl_pardiso->a, mat_mkl_pardiso->ia, mat_mkl_pardiso->ja, mat_mkl_pardiso->perm,
453: &mat_mkl_pardiso->nrhs, mat_mkl_pardiso->iparm, &mat_mkl_pardiso->msglvl, (void *)barray, (void *)xarray, &mat_mkl_pardiso->err));
454: PetscCheck(mat_mkl_pardiso->err >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error reported by MKL PARDISO: err=%" PetscInt_FMT ". Please check manual", (PetscInt)mat_mkl_pardiso->err);
456: PetscCall(VecRestoreArrayRead(b, &barray));
457: PetscCall(VecRestoreArrayWrite(x, &xarray));
458: mat_mkl_pardiso->CleanUp = PETSC_TRUE;
459: PetscFunctionReturn(PETSC_SUCCESS);
460: }
462: static PetscErrorCode MatBackwardSolve_MKL_PARDISO(Mat A, Vec b, Vec x)
463: {
464: Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)A->data;
465: PetscScalar *xarray;
466: const PetscScalar *barray;
468: PetscFunctionBegin;
469: PetscCheck(!mat_mkl_pardiso->schur, PETSC_COMM_SELF, PETSC_ERR_SUP, "Backward substitution not supported with Schur complement");
471: mat_mkl_pardiso->nrhs = 1;
472: PetscCall(VecGetArrayWrite(x, &xarray));
473: PetscCall(VecGetArrayRead(b, &barray));
475: mat_mkl_pardiso->phase = JOB_SOLVE_BACKWARD_SUBSTITUTION;
477: PetscCallPardiso(MKL_PARDISO(mat_mkl_pardiso->pt, &mat_mkl_pardiso->maxfct, &mat_mkl_pardiso->mnum, &mat_mkl_pardiso->mtype, &mat_mkl_pardiso->phase, &mat_mkl_pardiso->n, mat_mkl_pardiso->a, mat_mkl_pardiso->ia, mat_mkl_pardiso->ja, mat_mkl_pardiso->perm,
478: &mat_mkl_pardiso->nrhs, mat_mkl_pardiso->iparm, &mat_mkl_pardiso->msglvl, (void *)barray, (void *)xarray, &mat_mkl_pardiso->err));
479: PetscCheck(mat_mkl_pardiso->err >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error reported by MKL PARDISO: err=%" PetscInt_FMT ". Please check manual", (PetscInt)mat_mkl_pardiso->err);
481: PetscCall(VecRestoreArrayRead(b, &barray));
482: PetscCall(VecRestoreArrayWrite(x, &xarray));
483: mat_mkl_pardiso->CleanUp = PETSC_TRUE;
484: PetscFunctionReturn(PETSC_SUCCESS);
485: }
487: static PetscErrorCode MatSolveTranspose_MKL_PARDISO(Mat A, Vec b, Vec x)
488: {
489: Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)A->data;
490: PetscInt oiparm12;
492: PetscFunctionBegin;
493: oiparm12 = mat_mkl_pardiso->iparm[12 - 1];
494: mat_mkl_pardiso->iparm[12 - 1] = 2;
495: PetscCall(MatSolve_MKL_PARDISO(A, b, x));
496: mat_mkl_pardiso->iparm[12 - 1] = oiparm12;
497: PetscFunctionReturn(PETSC_SUCCESS);
498: }
500: static PetscErrorCode MatMatSolve_MKL_PARDISO(Mat A, Mat B, Mat X)
501: {
502: Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)A->data;
503: const PetscScalar *barray;
504: PetscScalar *xarray;
505: PetscBool flg;
507: PetscFunctionBegin;
508: PetscCall(PetscObjectBaseTypeCompare((PetscObject)B, MATSEQDENSE, &flg));
509: PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Matrix B must be MATSEQDENSE matrix");
510: if (X != B) {
511: PetscCall(PetscObjectBaseTypeCompare((PetscObject)X, MATSEQDENSE, &flg));
512: PetscCheck(flg, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Matrix X must be MATSEQDENSE matrix");
513: }
515: PetscCall(MatGetSize(B, NULL, (PetscInt *)&mat_mkl_pardiso->nrhs));
517: if (mat_mkl_pardiso->nrhs > 0) {
518: PetscCall(MatDenseGetArrayRead(B, &barray));
519: PetscCall(MatDenseGetArrayWrite(X, &xarray));
521: PetscCheck(barray != xarray, PETSC_COMM_SELF, PETSC_ERR_SUP, "B and X cannot share the same memory location");
522: if (!mat_mkl_pardiso->schur) mat_mkl_pardiso->phase = JOB_SOLVE_ITERATIVE_REFINEMENT;
523: else mat_mkl_pardiso->phase = JOB_SOLVE_FORWARD_SUBSTITUTION;
525: PetscCallPardiso(MKL_PARDISO(mat_mkl_pardiso->pt, &mat_mkl_pardiso->maxfct, &mat_mkl_pardiso->mnum, &mat_mkl_pardiso->mtype, &mat_mkl_pardiso->phase, &mat_mkl_pardiso->n, mat_mkl_pardiso->a, mat_mkl_pardiso->ia, mat_mkl_pardiso->ja,
526: mat_mkl_pardiso->perm, &mat_mkl_pardiso->nrhs, mat_mkl_pardiso->iparm, &mat_mkl_pardiso->msglvl, (void *)barray, (void *)xarray, &mat_mkl_pardiso->err));
527: PetscCheck(mat_mkl_pardiso->err >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error reported by MKL PARDISO: err=%" PetscInt_FMT ". Please check manual", (PetscInt)mat_mkl_pardiso->err);
529: PetscCall(MatDenseRestoreArrayRead(B, &barray));
530: if (mat_mkl_pardiso->schur) { /* solve Schur complement and expand solution */
531: PetscScalar *o_schur_work = NULL;
533: /* solve Schur complement */
534: if (!mat_mkl_pardiso->solve_interior) {
535: PetscInt shift = mat_mkl_pardiso->schur_size * mat_mkl_pardiso->nrhs, scale;
536: PetscInt mem = mat_mkl_pardiso->n * mat_mkl_pardiso->nrhs;
538: PetscCall(MatFactorFactorizeSchurComplement(A));
539: /* allocate extra memory if it is needed */
540: scale = 1;
541: if (A->schur_status == MAT_FACTOR_SCHUR_INVERTED) scale = 2;
542: mem *= scale;
543: if (mem > mat_mkl_pardiso->schur_work_size) {
544: o_schur_work = mat_mkl_pardiso->schur_work;
545: PetscCall(PetscMalloc1(mem, &mat_mkl_pardiso->schur_work));
546: }
547: /* if inverted, uses BLAS *MM subroutines, otherwise LAPACK *TRS */
548: if (A->schur_status != MAT_FACTOR_SCHUR_INVERTED) shift = 0;
549: PetscCall(MatMKLPardisoScatterSchur_Private(mat_mkl_pardiso, xarray, mat_mkl_pardiso->schur_work, PETSC_TRUE));
550: PetscCall(MatMKLPardisoSolveSchur_Private(A, mat_mkl_pardiso->schur_work, mat_mkl_pardiso->schur_work + shift));
551: PetscCall(MatMKLPardisoScatterSchur_Private(mat_mkl_pardiso, xarray, mat_mkl_pardiso->schur_work + shift, PETSC_FALSE));
552: } else { /* if we are solving for the interior problem, any value in barray[schur,n] forward-substituted to xarray[schur,n] will be neglected */
553: PetscInt i, n, m = 0;
554: for (n = 0; n < mat_mkl_pardiso->nrhs; n++) {
555: for (i = 0; i < mat_mkl_pardiso->schur_size; i++) xarray[mat_mkl_pardiso->schur_idxs[i] + m] = 0.;
556: m += mat_mkl_pardiso->n;
557: }
558: }
560: /* expansion phase */
561: mat_mkl_pardiso->iparm[6 - 1] = 1;
562: mat_mkl_pardiso->phase = JOB_SOLVE_BACKWARD_SUBSTITUTION;
563: PetscCallPardiso(MKL_PARDISO(mat_mkl_pardiso->pt, &mat_mkl_pardiso->maxfct, &mat_mkl_pardiso->mnum, &mat_mkl_pardiso->mtype, &mat_mkl_pardiso->phase, &mat_mkl_pardiso->n, mat_mkl_pardiso->a, mat_mkl_pardiso->ia, mat_mkl_pardiso->ja,
564: mat_mkl_pardiso->perm, &mat_mkl_pardiso->nrhs, mat_mkl_pardiso->iparm, &mat_mkl_pardiso->msglvl, (void *)xarray, (void *)mat_mkl_pardiso->schur_work, /* according to the specs, the solution vector is always used */
565: &mat_mkl_pardiso->err));
566: if (o_schur_work) { /* restore original Schur_work (minimal size) */
567: PetscCall(PetscFree(mat_mkl_pardiso->schur_work));
568: mat_mkl_pardiso->schur_work = o_schur_work;
569: }
570: PetscCheck(mat_mkl_pardiso->err >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error reported by MKL PARDISO: err=%" PetscInt_FMT ". Please check manual", (PetscInt)mat_mkl_pardiso->err);
571: mat_mkl_pardiso->iparm[6 - 1] = 0;
572: }
573: PetscCall(MatDenseRestoreArrayWrite(X, &xarray));
574: }
575: mat_mkl_pardiso->CleanUp = PETSC_TRUE;
576: PetscFunctionReturn(PETSC_SUCCESS);
577: }
579: static PetscErrorCode MatFactorNumeric_MKL_PARDISO(Mat F, Mat A, const MatFactorInfo *info)
580: {
581: Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)F->data;
583: PetscFunctionBegin;
584: mat_mkl_pardiso->matstruc = SAME_NONZERO_PATTERN;
585: PetscCall((*mat_mkl_pardiso->Convert)(A, mat_mkl_pardiso->needsym, MAT_REUSE_MATRIX, &mat_mkl_pardiso->freeaij, &mat_mkl_pardiso->nz, &mat_mkl_pardiso->ia, &mat_mkl_pardiso->ja, (PetscScalar **)&mat_mkl_pardiso->a));
587: mat_mkl_pardiso->phase = JOB_NUMERICAL_FACTORIZATION;
588: PetscCallPardiso(MKL_PARDISO(mat_mkl_pardiso->pt, &mat_mkl_pardiso->maxfct, &mat_mkl_pardiso->mnum, &mat_mkl_pardiso->mtype, &mat_mkl_pardiso->phase, &mat_mkl_pardiso->n, mat_mkl_pardiso->a, mat_mkl_pardiso->ia, mat_mkl_pardiso->ja, mat_mkl_pardiso->perm,
589: &mat_mkl_pardiso->nrhs, mat_mkl_pardiso->iparm, &mat_mkl_pardiso->msglvl, NULL, (void *)mat_mkl_pardiso->schur, &mat_mkl_pardiso->err));
590: PetscCheck(mat_mkl_pardiso->err >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error reported by MKL PARDISO: err=%" PetscInt_FMT ". Please check manual", (PetscInt)mat_mkl_pardiso->err);
592: /* report flops */
593: if (mat_mkl_pardiso->iparm[18] > 0) PetscCall(PetscLogFlops(PetscPowRealInt(10., 6) * mat_mkl_pardiso->iparm[18]));
595: if (F->schur) { /* schur output from pardiso is in row major format */
596: #if defined(PETSC_HAVE_CUDA)
597: F->schur->offloadmask = PETSC_OFFLOAD_CPU;
598: #endif
599: PetscCall(MatFactorRestoreSchurComplement(F, NULL, MAT_FACTOR_SCHUR_UNFACTORED));
600: PetscCall(MatTranspose(F->schur, MAT_INPLACE_MATRIX, &F->schur));
601: }
602: mat_mkl_pardiso->matstruc = SAME_NONZERO_PATTERN;
603: mat_mkl_pardiso->CleanUp = PETSC_TRUE;
604: PetscFunctionReturn(PETSC_SUCCESS);
605: }
607: static PetscErrorCode MatSetFromOptions_MKL_PARDISO(Mat F, Mat A)
608: {
609: Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)F->data;
610: PetscInt icntl, bs, threads = 1;
611: PetscBool flg;
613: PetscFunctionBegin;
614: PetscOptionsBegin(PetscObjectComm((PetscObject)F), ((PetscObject)F)->prefix, "MKL_PARDISO Options", "Mat");
616: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_65", "Suggested number of threads to use within MKL PARDISO", "None", threads, &threads, &flg));
617: if (flg) PetscSetMKL_PARDISOThreads((int)threads);
619: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_66", "Maximum number of factors with identical sparsity structure that must be kept in memory at the same time", "None", mat_mkl_pardiso->maxfct, &icntl, &flg));
620: if (flg) mat_mkl_pardiso->maxfct = icntl;
622: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_67", "Indicates the actual matrix for the solution phase", "None", mat_mkl_pardiso->mnum, &icntl, &flg));
623: if (flg) mat_mkl_pardiso->mnum = icntl;
625: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_68", "Message level information", "None", mat_mkl_pardiso->msglvl, &icntl, &flg));
626: if (flg) mat_mkl_pardiso->msglvl = icntl;
628: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_69", "Defines the matrix type", "None", mat_mkl_pardiso->mtype, &icntl, &flg));
629: if (flg) {
630: void *pt[IPARM_SIZE];
631: mat_mkl_pardiso->mtype = icntl;
632: icntl = mat_mkl_pardiso->iparm[34];
633: bs = mat_mkl_pardiso->iparm[36];
634: MKL_PARDISO_INIT(pt, &mat_mkl_pardiso->mtype, mat_mkl_pardiso->iparm);
635: #if defined(PETSC_USE_REAL_SINGLE)
636: mat_mkl_pardiso->iparm[27] = 1;
637: #else
638: mat_mkl_pardiso->iparm[27] = 0;
639: #endif
640: mat_mkl_pardiso->iparm[34] = icntl;
641: mat_mkl_pardiso->iparm[36] = bs;
642: }
644: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_1", "Use default values (if 0)", "None", mat_mkl_pardiso->iparm[0], &icntl, &flg));
645: if (flg) mat_mkl_pardiso->iparm[0] = icntl;
647: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_2", "Fill-in reducing ordering for the input matrix", "None", mat_mkl_pardiso->iparm[1], &icntl, &flg));
648: if (flg) mat_mkl_pardiso->iparm[1] = icntl;
650: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_4", "Preconditioned CGS/CG", "None", mat_mkl_pardiso->iparm[3], &icntl, &flg));
651: if (flg) mat_mkl_pardiso->iparm[3] = icntl;
653: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_5", "User permutation", "None", mat_mkl_pardiso->iparm[4], &icntl, &flg));
654: if (flg) mat_mkl_pardiso->iparm[4] = icntl;
656: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_6", "Write solution on x", "None", mat_mkl_pardiso->iparm[5], &icntl, &flg));
657: if (flg) mat_mkl_pardiso->iparm[5] = icntl;
659: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_8", "Iterative refinement step", "None", mat_mkl_pardiso->iparm[7], &icntl, &flg));
660: if (flg) mat_mkl_pardiso->iparm[7] = icntl;
662: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_10", "Pivoting perturbation", "None", mat_mkl_pardiso->iparm[9], &icntl, &flg));
663: if (flg) mat_mkl_pardiso->iparm[9] = icntl;
665: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_11", "Scaling vectors", "None", mat_mkl_pardiso->iparm[10], &icntl, &flg));
666: if (flg) mat_mkl_pardiso->iparm[10] = icntl;
668: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_12", "Solve with transposed or conjugate transposed matrix A", "None", mat_mkl_pardiso->iparm[11], &icntl, &flg));
669: if (flg) mat_mkl_pardiso->iparm[11] = icntl;
671: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_13", "Improved accuracy using (non-) symmetric weighted matching", "None", mat_mkl_pardiso->iparm[12], &icntl, &flg));
672: if (flg) mat_mkl_pardiso->iparm[12] = icntl;
674: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_18", "Numbers of non-zero elements", "None", mat_mkl_pardiso->iparm[17], &icntl, &flg));
675: if (flg) mat_mkl_pardiso->iparm[17] = icntl;
677: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_19", "Report number of floating point operations (0 to disable)", "None", mat_mkl_pardiso->iparm[18], &icntl, &flg));
678: if (flg) mat_mkl_pardiso->iparm[18] = icntl;
680: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_21", "Pivoting for symmetric indefinite matrices", "None", mat_mkl_pardiso->iparm[20], &icntl, &flg));
681: if (flg) mat_mkl_pardiso->iparm[20] = icntl;
683: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_24", "Parallel factorization control", "None", mat_mkl_pardiso->iparm[23], &icntl, &flg));
684: if (flg) mat_mkl_pardiso->iparm[23] = icntl;
686: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_25", "Parallel forward/backward solve control", "None", mat_mkl_pardiso->iparm[24], &icntl, &flg));
687: if (flg) mat_mkl_pardiso->iparm[24] = icntl;
689: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_27", "Matrix checker", "None", mat_mkl_pardiso->iparm[26], &icntl, &flg));
690: if (flg) mat_mkl_pardiso->iparm[26] = icntl;
692: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_31", "Partial solve and computing selected components of the solution vectors", "None", mat_mkl_pardiso->iparm[30], &icntl, &flg));
693: if (flg) mat_mkl_pardiso->iparm[30] = icntl;
695: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_34", "Optimal number of threads for conditional numerical reproducibility (CNR) mode", "None", mat_mkl_pardiso->iparm[33], &icntl, &flg));
696: if (flg) mat_mkl_pardiso->iparm[33] = icntl;
698: PetscCall(PetscOptionsInt("-mat_mkl_pardiso_60", "Intel MKL PARDISO mode", "None", mat_mkl_pardiso->iparm[59], &icntl, &flg));
699: if (flg) mat_mkl_pardiso->iparm[59] = icntl;
700: PetscOptionsEnd();
701: PetscFunctionReturn(PETSC_SUCCESS);
702: }
704: static PetscErrorCode MatFactorMKL_PARDISOInitialize_Private(Mat A, MatFactorType ftype, Mat_MKL_PARDISO *mat_mkl_pardiso)
705: {
706: PetscInt i, bs;
707: PetscBool match;
709: PetscFunctionBegin;
710: for (i = 0; i < IPARM_SIZE; i++) mat_mkl_pardiso->iparm[i] = 0;
711: for (i = 0; i < IPARM_SIZE; i++) mat_mkl_pardiso->pt[i] = 0;
712: #if defined(PETSC_USE_REAL_SINGLE)
713: mat_mkl_pardiso->iparm[27] = 1;
714: #else
715: mat_mkl_pardiso->iparm[27] = 0;
716: #endif
717: /* Default options for both sym and unsym */
718: mat_mkl_pardiso->iparm[0] = 1; /* Solver default parameters overridden with provided by iparm */
719: mat_mkl_pardiso->iparm[1] = 2; /* Metis reordering */
720: mat_mkl_pardiso->iparm[5] = 0; /* Write solution into x */
721: mat_mkl_pardiso->iparm[7] = 0; /* Max number of iterative refinement steps */
722: mat_mkl_pardiso->iparm[17] = -1; /* Output: Number of nonzeros in the factor LU */
723: mat_mkl_pardiso->iparm[18] = -1; /* Output: Mflops for LU factorization */
724: #if 0
725: mat_mkl_pardiso->iparm[23] = 1; /* Parallel factorization control*/
726: #endif
727: PetscCall(PetscObjectTypeCompareAny((PetscObject)A, &match, MATSEQBAIJ, MATSEQSBAIJ, ""));
728: PetscCall(MatGetBlockSize(A, &bs));
729: if (!match || bs == 1) {
730: mat_mkl_pardiso->iparm[34] = 1; /* Cluster Sparse Solver use C-style indexing for ia and ja arrays */
731: mat_mkl_pardiso->n = A->rmap->N;
732: } else {
733: mat_mkl_pardiso->iparm[34] = 0; /* Cluster Sparse Solver use Fortran-style indexing for ia and ja arrays */
734: mat_mkl_pardiso->iparm[36] = bs;
735: mat_mkl_pardiso->n = A->rmap->N / bs;
736: }
737: mat_mkl_pardiso->iparm[39] = 0; /* Input: matrix/rhs/solution stored on rank-0 */
739: mat_mkl_pardiso->CleanUp = PETSC_FALSE;
740: mat_mkl_pardiso->maxfct = 1; /* Maximum number of numerical factorizations. */
741: mat_mkl_pardiso->mnum = 1; /* Which factorization to use. */
742: mat_mkl_pardiso->msglvl = 0; /* 0: do not print 1: Print statistical information in file */
743: mat_mkl_pardiso->phase = -1;
744: mat_mkl_pardiso->err = 0;
746: mat_mkl_pardiso->nrhs = 1;
747: mat_mkl_pardiso->err = 0;
748: mat_mkl_pardiso->phase = -1;
750: if (ftype == MAT_FACTOR_LU) {
751: mat_mkl_pardiso->iparm[9] = 13; /* Perturb the pivot elements with 1E-13 */
752: mat_mkl_pardiso->iparm[10] = 1; /* Use nonsymmetric permutation and scaling MPS */
753: mat_mkl_pardiso->iparm[12] = 1; /* Switch on Maximum Weighted Matching algorithm (default for non-symmetric) */
754: } else {
755: mat_mkl_pardiso->iparm[9] = 8; /* Perturb the pivot elements with 1E-8 */
756: mat_mkl_pardiso->iparm[10] = 0; /* Use nonsymmetric permutation and scaling MPS */
757: mat_mkl_pardiso->iparm[12] = 1; /* Switch on Maximum Weighted Matching algorithm (default for non-symmetric) */
758: #if defined(PETSC_USE_DEBUG)
759: mat_mkl_pardiso->iparm[26] = 1; /* Matrix checker */
760: #endif
761: }
762: PetscCall(PetscCalloc1(A->rmap->N * sizeof(INT_TYPE), &mat_mkl_pardiso->perm));
763: mat_mkl_pardiso->schur_size = 0;
764: PetscFunctionReturn(PETSC_SUCCESS);
765: }
767: static PetscErrorCode MatFactorSymbolic_AIJMKL_PARDISO_Private(Mat F, Mat A, const MatFactorInfo *info)
768: {
769: Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)F->data;
771: PetscFunctionBegin;
772: mat_mkl_pardiso->matstruc = DIFFERENT_NONZERO_PATTERN;
773: PetscCall(MatSetFromOptions_MKL_PARDISO(F, A));
774: /* throw away any previously computed structure */
775: if (mat_mkl_pardiso->freeaij) {
776: PetscCall(PetscFree2(mat_mkl_pardiso->ia, mat_mkl_pardiso->ja));
777: if (mat_mkl_pardiso->iparm[34] == 1) PetscCall(PetscFree(mat_mkl_pardiso->a));
778: }
779: PetscCall((*mat_mkl_pardiso->Convert)(A, mat_mkl_pardiso->needsym, MAT_INITIAL_MATRIX, &mat_mkl_pardiso->freeaij, &mat_mkl_pardiso->nz, &mat_mkl_pardiso->ia, &mat_mkl_pardiso->ja, (PetscScalar **)&mat_mkl_pardiso->a));
780: if (mat_mkl_pardiso->iparm[34] == 1) mat_mkl_pardiso->n = A->rmap->N;
781: else mat_mkl_pardiso->n = A->rmap->N / A->rmap->bs;
783: mat_mkl_pardiso->phase = JOB_ANALYSIS;
785: /* reset flops counting if requested */
786: if (mat_mkl_pardiso->iparm[18]) mat_mkl_pardiso->iparm[18] = -1;
788: PetscCallPardiso(MKL_PARDISO(mat_mkl_pardiso->pt, &mat_mkl_pardiso->maxfct, &mat_mkl_pardiso->mnum, &mat_mkl_pardiso->mtype, &mat_mkl_pardiso->phase, &mat_mkl_pardiso->n, mat_mkl_pardiso->a, mat_mkl_pardiso->ia, mat_mkl_pardiso->ja, mat_mkl_pardiso->perm,
789: &mat_mkl_pardiso->nrhs, mat_mkl_pardiso->iparm, &mat_mkl_pardiso->msglvl, NULL, NULL, &mat_mkl_pardiso->err));
790: PetscCheck(mat_mkl_pardiso->err >= 0, PETSC_COMM_SELF, PETSC_ERR_LIB, "Error reported by MKL PARDISO: err=%" PetscInt_FMT ". Please check manual", (PetscInt)mat_mkl_pardiso->err);
792: mat_mkl_pardiso->CleanUp = PETSC_TRUE;
794: if (F->factortype == MAT_FACTOR_LU) F->ops->lufactornumeric = MatFactorNumeric_MKL_PARDISO;
795: else F->ops->choleskyfactornumeric = MatFactorNumeric_MKL_PARDISO;
797: F->ops->solve = MatSolve_MKL_PARDISO;
798: F->ops->solvetranspose = MatSolveTranspose_MKL_PARDISO;
799: F->ops->matsolve = MatMatSolve_MKL_PARDISO;
800: if (F->factortype == MAT_FACTOR_LU || (!PetscDefined(USE_COMPLEX) && F->factortype == MAT_FACTOR_CHOLESKY && A->spd == PETSC_BOOL3_TRUE)) {
801: F->ops->backwardsolve = MatBackwardSolve_MKL_PARDISO;
802: F->ops->forwardsolve = MatForwardSolve_MKL_PARDISO;
803: }
804: PetscFunctionReturn(PETSC_SUCCESS);
805: }
807: static PetscErrorCode MatLUFactorSymbolic_AIJMKL_PARDISO(Mat F, Mat A, IS r, IS c, const MatFactorInfo *info)
808: {
809: PetscFunctionBegin;
810: PetscCall(MatFactorSymbolic_AIJMKL_PARDISO_Private(F, A, info));
811: PetscFunctionReturn(PETSC_SUCCESS);
812: }
814: #if !defined(PETSC_USE_COMPLEX)
815: static PetscErrorCode MatGetInertia_MKL_PARDISO(Mat F, PetscInt *nneg, PetscInt *nzero, PetscInt *npos)
816: {
817: Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)F->data;
819: PetscFunctionBegin;
820: if (nneg) *nneg = mat_mkl_pardiso->iparm[22];
821: if (npos) *npos = mat_mkl_pardiso->iparm[21];
822: if (nzero) *nzero = F->rmap->N - (mat_mkl_pardiso->iparm[22] + mat_mkl_pardiso->iparm[21]);
823: PetscFunctionReturn(PETSC_SUCCESS);
824: }
825: #endif
827: static PetscErrorCode MatCholeskyFactorSymbolic_AIJMKL_PARDISO(Mat F, Mat A, IS r, const MatFactorInfo *info)
828: {
829: PetscFunctionBegin;
830: PetscCall(MatFactorSymbolic_AIJMKL_PARDISO_Private(F, A, info));
831: F->ops->getinertia = NULL;
832: #if !defined(PETSC_USE_COMPLEX)
833: F->ops->getinertia = MatGetInertia_MKL_PARDISO;
834: #endif
835: PetscFunctionReturn(PETSC_SUCCESS);
836: }
838: static PetscErrorCode MatView_MKL_PARDISO(Mat A, PetscViewer viewer)
839: {
840: PetscBool iascii;
841: PetscViewerFormat format;
842: Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)A->data;
843: PetscInt i;
845: PetscFunctionBegin;
846: if (A->ops->solve != MatSolve_MKL_PARDISO) PetscFunctionReturn(PETSC_SUCCESS);
848: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
849: if (iascii) {
850: PetscCall(PetscViewerGetFormat(viewer, &format));
851: if (format == PETSC_VIEWER_ASCII_INFO) {
852: PetscCall(PetscViewerASCIIPrintf(viewer, "MKL PARDISO run parameters:\n"));
853: PetscCall(PetscViewerASCIIPrintf(viewer, "MKL PARDISO phase: %" PetscInt_FMT "\n", (PetscInt)mat_mkl_pardiso->phase));
854: for (i = 1; i <= 64; i++) PetscCall(PetscViewerASCIIPrintf(viewer, "MKL PARDISO iparm[%" PetscInt_FMT "]: %" PetscInt_FMT "\n", i, (PetscInt)mat_mkl_pardiso->iparm[i - 1]));
855: PetscCall(PetscViewerASCIIPrintf(viewer, "MKL PARDISO maxfct: %" PetscInt_FMT "\n", (PetscInt)mat_mkl_pardiso->maxfct));
856: PetscCall(PetscViewerASCIIPrintf(viewer, "MKL PARDISO mnum: %" PetscInt_FMT "\n", (PetscInt)mat_mkl_pardiso->mnum));
857: PetscCall(PetscViewerASCIIPrintf(viewer, "MKL PARDISO mtype: %" PetscInt_FMT "\n", (PetscInt)mat_mkl_pardiso->mtype));
858: PetscCall(PetscViewerASCIIPrintf(viewer, "MKL PARDISO n: %" PetscInt_FMT "\n", (PetscInt)mat_mkl_pardiso->n));
859: PetscCall(PetscViewerASCIIPrintf(viewer, "MKL PARDISO nrhs: %" PetscInt_FMT "\n", (PetscInt)mat_mkl_pardiso->nrhs));
860: PetscCall(PetscViewerASCIIPrintf(viewer, "MKL PARDISO msglvl: %" PetscInt_FMT "\n", (PetscInt)mat_mkl_pardiso->msglvl));
861: }
862: }
863: PetscFunctionReturn(PETSC_SUCCESS);
864: }
866: static PetscErrorCode MatGetInfo_MKL_PARDISO(Mat A, MatInfoType flag, MatInfo *info)
867: {
868: Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)A->data;
870: PetscFunctionBegin;
871: info->block_size = 1.0;
872: info->nz_used = mat_mkl_pardiso->iparm[17];
873: info->nz_allocated = mat_mkl_pardiso->iparm[17];
874: info->nz_unneeded = 0.0;
875: info->assemblies = 0.0;
876: info->mallocs = 0.0;
877: info->memory = 0.0;
878: info->fill_ratio_given = 0;
879: info->fill_ratio_needed = 0;
880: info->factor_mallocs = 0;
881: PetscFunctionReturn(PETSC_SUCCESS);
882: }
884: static PetscErrorCode MatMkl_PardisoSetCntl_MKL_PARDISO(Mat F, PetscInt icntl, PetscInt ival)
885: {
886: PetscInt backup, bs;
887: Mat_MKL_PARDISO *mat_mkl_pardiso = (Mat_MKL_PARDISO *)F->data;
889: PetscFunctionBegin;
890: if (icntl <= 64) {
891: mat_mkl_pardiso->iparm[icntl - 1] = ival;
892: } else {
893: if (icntl == 65) PetscSetMKL_PARDISOThreads((int)ival);
894: else if (icntl == 66) mat_mkl_pardiso->maxfct = ival;
895: else if (icntl == 67) mat_mkl_pardiso->mnum = ival;
896: else if (icntl == 68) mat_mkl_pardiso->msglvl = ival;
897: else if (icntl == 69) {
898: void *pt[IPARM_SIZE];
899: backup = mat_mkl_pardiso->iparm[34];
900: bs = mat_mkl_pardiso->iparm[36];
901: mat_mkl_pardiso->mtype = ival;
902: MKL_PARDISO_INIT(pt, &mat_mkl_pardiso->mtype, mat_mkl_pardiso->iparm);
903: #if defined(PETSC_USE_REAL_SINGLE)
904: mat_mkl_pardiso->iparm[27] = 1;
905: #else
906: mat_mkl_pardiso->iparm[27] = 0;
907: #endif
908: mat_mkl_pardiso->iparm[34] = backup;
909: mat_mkl_pardiso->iparm[36] = bs;
910: } else if (icntl == 70) mat_mkl_pardiso->solve_interior = (PetscBool)!!ival;
911: }
912: PetscFunctionReturn(PETSC_SUCCESS);
913: }
915: /*@
916: MatMkl_PardisoSetCntl - Set MKL PARDISO <https://www.intel.com/content/www/us/en/docs/onemkl/developer-reference-c/2023-2/onemkl-pardiso-parallel-direct-sparse-solver-iface.html> parameters
918: Logically Collective
920: Input Parameters:
921: + F - the factored matrix obtained by calling `MatGetFactor()`
922: . icntl - index of MKL PARDISO parameter
923: - ival - value of MKL PARDISO parameter
925: Options Database Key:
926: . -mat_mkl_pardiso_<icntl> <ival> - change the option numbered icntl to the value ival
928: Level: beginner
930: .seealso: [](ch_matrices), `Mat`, `MATSOLVERMKL_PARDISO`, `MatGetFactor()`
931: @*/
932: PetscErrorCode MatMkl_PardisoSetCntl(Mat F, PetscInt icntl, PetscInt ival)
933: {
934: PetscFunctionBegin;
935: PetscTryMethod(F, "MatMkl_PardisoSetCntl_C", (Mat, PetscInt, PetscInt), (F, icntl, ival));
936: PetscFunctionReturn(PETSC_SUCCESS);
937: }
939: /*MC
940: MATSOLVERMKL_PARDISO - A matrix type providing direct solvers, LU, for
941: `MATSEQAIJ` matrices via the external package MKL PARDISO
942: <https://www.intel.com/content/www/us/en/docs/onemkl/developer-reference-c/2024-0/onemkl-pardiso-parallel-direct-sparse-solver-iface.html>.
944: Use `-pc_type lu` `-pc_factor_mat_solver_type mkl_pardiso` to use this direct solver
946: Options Database Keys:
947: + -mat_mkl_pardiso_65 - Suggested number of threads to use within MKL PARDISO
948: . -mat_mkl_pardiso_66 - Maximum number of factors with identical sparsity structure that must be kept in memory at the same time
949: . -mat_mkl_pardiso_67 - Indicates the actual matrix for the solution phase
950: . -mat_mkl_pardiso_68 - Message level information, use 1 to get detailed information on the solver options
951: . -mat_mkl_pardiso_69 - Defines the matrix type. IMPORTANT: When you set this flag, iparm parameters are going to be set to the default ones for the matrix type
952: . -mat_mkl_pardiso_1 - Use default values
953: . -mat_mkl_pardiso_2 - Fill-in reducing ordering for the input matrix
954: . -mat_mkl_pardiso_4 - Preconditioned CGS/CG
955: . -mat_mkl_pardiso_5 - User permutation
956: . -mat_mkl_pardiso_6 - Write solution on x
957: . -mat_mkl_pardiso_8 - Iterative refinement step
958: . -mat_mkl_pardiso_10 - Pivoting perturbation
959: . -mat_mkl_pardiso_11 - Scaling vectors
960: . -mat_mkl_pardiso_12 - Solve with transposed or conjugate transposed matrix A
961: . -mat_mkl_pardiso_13 - Improved accuracy using (non-) symmetric weighted matching
962: . -mat_mkl_pardiso_18 - Numbers of non-zero elements
963: . -mat_mkl_pardiso_19 - Report number of floating point operations
964: . -mat_mkl_pardiso_21 - Pivoting for symmetric indefinite matrices
965: . -mat_mkl_pardiso_24 - Parallel factorization control
966: . -mat_mkl_pardiso_25 - Parallel forward/backward solve control
967: . -mat_mkl_pardiso_27 - Matrix checker
968: . -mat_mkl_pardiso_31 - Partial solve and computing selected components of the solution vectors
969: . -mat_mkl_pardiso_34 - Optimal number of threads for conditional numerical reproducibility (CNR) mode
970: - -mat_mkl_pardiso_60 - Intel MKL PARDISO mode
972: Level: beginner
974: Notes:
975: Use `-mat_mkl_pardiso_68 1` to display the number of threads the solver is using. MKL does not provide a way to directly access this
976: information.
978: For more information on the options check the MKL PARDISO manual
980: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ`, `PCFactorSetMatSolverType()`, `MatSolverType`, `MatMkl_PardisoSetCntl()`, `MATSOLVERMKL_CPARDISO`
981: M*/
982: static PetscErrorCode MatFactorGetSolverType_mkl_pardiso(Mat A, MatSolverType *type)
983: {
984: PetscFunctionBegin;
985: *type = MATSOLVERMKL_PARDISO;
986: PetscFunctionReturn(PETSC_SUCCESS);
987: }
989: PETSC_EXTERN PetscErrorCode MatGetFactor_aij_mkl_pardiso(Mat A, MatFactorType ftype, Mat *F)
990: {
991: Mat B;
992: Mat_MKL_PARDISO *mat_mkl_pardiso;
993: PetscBool isSeqAIJ, isSeqBAIJ, isSeqSBAIJ;
995: PetscFunctionBegin;
996: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATSEQAIJ, &isSeqAIJ));
997: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQBAIJ, &isSeqBAIJ));
998: PetscCall(PetscObjectTypeCompare((PetscObject)A, MATSEQSBAIJ, &isSeqSBAIJ));
999: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &B));
1000: PetscCall(MatSetSizes(B, A->rmap->n, A->cmap->n, A->rmap->N, A->cmap->N));
1001: PetscCall(PetscStrallocpy("mkl_pardiso", &((PetscObject)B)->type_name));
1002: PetscCall(MatSetUp(B));
1004: PetscCall(PetscNew(&mat_mkl_pardiso));
1005: B->data = mat_mkl_pardiso;
1007: PetscCall(MatFactorMKL_PARDISOInitialize_Private(A, ftype, mat_mkl_pardiso));
1008: if (ftype == MAT_FACTOR_LU) {
1009: B->ops->lufactorsymbolic = MatLUFactorSymbolic_AIJMKL_PARDISO;
1010: B->factortype = MAT_FACTOR_LU;
1011: mat_mkl_pardiso->needsym = PETSC_FALSE;
1012: if (isSeqAIJ) mat_mkl_pardiso->Convert = MatMKLPardiso_Convert_seqaij;
1013: else if (isSeqBAIJ) mat_mkl_pardiso->Convert = MatMKLPardiso_Convert_seqbaij;
1014: else {
1015: PetscCheck(!isSeqSBAIJ, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "No support for MKL PARDISO LU factor with SEQSBAIJ format! Use MAT_FACTOR_CHOLESKY instead");
1016: SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "No support for MKL PARDISO LU with %s format", ((PetscObject)A)->type_name);
1017: }
1018: #if defined(PETSC_USE_COMPLEX)
1019: mat_mkl_pardiso->mtype = 13;
1020: #else
1021: mat_mkl_pardiso->mtype = 11;
1022: #endif
1023: } else {
1024: B->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_AIJMKL_PARDISO;
1025: B->factortype = MAT_FACTOR_CHOLESKY;
1026: if (isSeqAIJ) mat_mkl_pardiso->Convert = MatMKLPardiso_Convert_seqaij;
1027: else if (isSeqBAIJ) mat_mkl_pardiso->Convert = MatMKLPardiso_Convert_seqbaij;
1028: else if (isSeqSBAIJ) mat_mkl_pardiso->Convert = MatMKLPardiso_Convert_seqsbaij;
1029: else SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "No support for PARDISO CHOLESKY with %s format", ((PetscObject)A)->type_name);
1031: mat_mkl_pardiso->needsym = PETSC_TRUE;
1032: #if !defined(PETSC_USE_COMPLEX)
1033: if (A->spd == PETSC_BOOL3_TRUE) mat_mkl_pardiso->mtype = 2;
1034: else mat_mkl_pardiso->mtype = -2;
1035: #else
1036: mat_mkl_pardiso->mtype = 6;
1037: PetscCheck(A->hermitian != PETSC_BOOL3_TRUE, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "No support for MKL PARDISO CHOLESKY with Hermitian matrices! Use MAT_FACTOR_LU instead");
1038: #endif
1039: }
1040: B->ops->destroy = MatDestroy_MKL_PARDISO;
1041: B->ops->view = MatView_MKL_PARDISO;
1042: B->ops->getinfo = MatGetInfo_MKL_PARDISO;
1043: B->factortype = ftype;
1044: B->assembled = PETSC_TRUE;
1046: PetscCall(PetscFree(B->solvertype));
1047: PetscCall(PetscStrallocpy(MATSOLVERMKL_PARDISO, &B->solvertype));
1049: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorGetSolverType_C", MatFactorGetSolverType_mkl_pardiso));
1050: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatFactorSetSchurIS_C", MatFactorSetSchurIS_MKL_PARDISO));
1051: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatMkl_PardisoSetCntl_C", MatMkl_PardisoSetCntl_MKL_PARDISO));
1053: *F = B;
1054: PetscFunctionReturn(PETSC_SUCCESS);
1055: }
1057: PETSC_INTERN PetscErrorCode MatSolverTypeRegister_MKL_Pardiso(void)
1058: {
1059: PetscFunctionBegin;
1060: PetscCall(MatSolverTypeRegister(MATSOLVERMKL_PARDISO, MATSEQAIJ, MAT_FACTOR_LU, MatGetFactor_aij_mkl_pardiso));
1061: PetscCall(MatSolverTypeRegister(MATSOLVERMKL_PARDISO, MATSEQAIJ, MAT_FACTOR_CHOLESKY, MatGetFactor_aij_mkl_pardiso));
1062: PetscCall(MatSolverTypeRegister(MATSOLVERMKL_PARDISO, MATSEQBAIJ, MAT_FACTOR_LU, MatGetFactor_aij_mkl_pardiso));
1063: PetscCall(MatSolverTypeRegister(MATSOLVERMKL_PARDISO, MATSEQSBAIJ, MAT_FACTOR_CHOLESKY, MatGetFactor_aij_mkl_pardiso));
1064: PetscFunctionReturn(PETSC_SUCCESS);
1065: }