seq/aijmkl/aijmkl.c

10 #include <../src/mat/impls/aij/seq/aijmkl/aijmkl.h>
33   Mat_SeqAIJMKL *aijmkl = (Mat_SeqAIJMKL *)A->spptr;
60   if (reuse == MAT_INITIAL_MATRIX) aijmkl = (Mat_SeqAIJMKL *)B->spptr;
62   if (aijmkl->sparse_optimized) PetscCallExternal(mkl_sparse_destroy, aijmkl->csrA);
75   Mat_SeqAIJMKL *aijmkl = (Mat_SeqAIJMKL *)A->spptr;
79   if (aijmkl) {
82     if (aijmkl->sparse_optimized) PetscCallExternal(mkl_sparse_destroy, aijmkl->csrA);
100  * If called with an AIJMKL matrix for which aijmkl->sparse_optimized == PETSC_TRUE, then it destroys the old matrix
115   Mat_SeqAIJMKL *aijmkl = (Mat_SeqAIJMKL *)A->spptr;
126   if (aijmkl->no_SpMV2) PetscFunctionReturn(PETSC_SUCCESS);
129   if (aijmkl->sparse_optimized) {
132     PetscCallExternal(mkl_sparse_destroy, aijmkl->csrA);
134   aijmkl->sparse_optimized = PETSC_FALSE;
137   aijmkl->descr.type = SPARSE_MATRIX_TYPE_GENERAL;
138   aijmkl->descr.mode = SPARSE_FILL_MODE_LOWER;
139   aijmkl->descr.diag = SPARSE_DIAG_NON_UNIT;
148     PetscCallExternal(mkl_sparse_x_create_csr, &aijmkl->csrA, SPARSE_INDEX_BASE_ZERO, (MKL_INT)m, (MKL_INT)n, (MKL_INT *)ai, (MKL_INT *)(ai + 1), (MKL_INT *)aj, aa);
149     PetscCallExternal(mkl_sparse_set_mv_hint, aijmkl->csrA, SPARSE_OPERATION_NON_TRANSPOSE, aijmkl->descr, 1000);
150     PetscCallExternal(mkl_sparse_set_memory_hint, aijmkl->csrA, SPARSE_MEMORY_AGGRESSIVE);
151     if (!aijmkl->no_SpMV2) PetscCallExternal(mkl_sparse_optimize, aijmkl->csrA);
152     aijmkl->sparse_optimized = PETSC_TRUE;
153     PetscCall(PetscObjectStateGet((PetscObject)A, &aijmkl->state));
155     aijmkl->csrA = NULL;
169   Mat_SeqAIJMKL      *aijmkl;
200   aijmkl       = (Mat_SeqAIJMKL *)A->spptr;
201   aijmkl->csrA = csrA;
206   aijmkl->descr.type = SPARSE_MATRIX_TYPE_GENERAL;
207   aijmkl->descr.mode = SPARSE_FILL_MODE_LOWER;
208   aijmkl->descr.diag = SPARSE_DIAG_NON_UNIT;
210     PetscCallExternal(mkl_sparse_set_mv_hint, aijmkl->csrA, SPARSE_OPERATION_NON_TRANSPOSE, aijmkl->descr, 1000);
211     PetscCallExternal(mkl_sparse_set_memory_hint, aijmkl->csrA, SPARSE_MEMORY_AGGRESSIVE);
213   PetscCall(PetscObjectStateGet((PetscObject)A, &aijmkl->state));
229   Mat_SeqAIJMKL      *aijmkl = (Mat_SeqAIJMKL *)A->spptr;
234   if (!aijmkl->csrA) PetscFunctionReturn(PETSC_SUCCESS);
237   PetscCallExternal(mkl_sparse_x_export_csr, aijmkl->csrA, &indexing, (MKL_INT *)&nrows, (MKL_INT *)&ncols, (MKL_INT **)&ai, (MKL_INT **)&unused, (MKL_INT **)&aj, &aa);
249   PetscCall(PetscObjectStateGet((PetscObject)A, &aijmkl->state));
253   aijmkl->sparse_optimized = PETSC_FALSE;
266   Mat_SeqAIJMKL      *aijmkl = (Mat_SeqAIJMKL *)A->spptr;
273   if (!aijmkl->csrA) {
279   PetscCallExternal(mkl_sparse_x_export_csr, aijmkl->csrA, &indexing, (MKL_INT *)&nrows, (MKL_INT *)&ncols, (MKL_INT **)&ai, (MKL_INT **)&unused, (MKL_INT **)&aj, &aa);
301   Mat_SeqAIJMKL *aijmkl = (Mat_SeqAIJMKL *)A->spptr;
307   PetscCall(PetscArraycpy(aijmkl_dest, aijmkl, 1));
309   if (aijmkl->eager_inspection) PetscCall(MatSeqAIJMKL_create_mkl_handle(A));
316   Mat_SeqAIJMKL *aijmkl;
331   aijmkl = (Mat_SeqAIJMKL *)A->spptr;
332   if (aijmkl->eager_inspection) PetscCall(MatSeqAIJMKL_create_mkl_handle(A));
376   Mat_SeqAIJMKL     *aijmkl = (Mat_SeqAIJMKL *)A->spptr;
397   if (!aijmkl->sparse_optimized || aijmkl->state != state) PetscCall(MatSeqAIJMKL_create_mkl_handle(A));
400   PetscCallExternal(mkl_sparse_x_mv, SPARSE_OPERATION_NON_TRANSPOSE, 1.0, aijmkl->csrA, aijmkl->descr, x, 0.0, y);
449   Mat_SeqAIJMKL     *aijmkl = (Mat_SeqAIJMKL *)A->spptr;
470   if (!aijmkl->sparse_optimized || aijmkl->state != state) PetscCall(MatSeqAIJMKL_create_mkl_handle(A));
473   PetscCallExternal(mkl_sparse_x_mv, SPARSE_OPERATION_TRANSPOSE, 1.0, aijmkl->csrA, aijmkl->descr, x, 0.0, y);
534   Mat_SeqAIJMKL     *aijmkl = (Mat_SeqAIJMKL *)A->spptr;
559   if (!aijmkl->sparse_optimized || aijmkl->state != state) PetscCall(MatSeqAIJMKL_create_mkl_handle(A));
565     PetscCallExternal(mkl_sparse_x_mv, SPARSE_OPERATION_NON_TRANSPOSE, 1.0, aijmkl->csrA, aijmkl->descr, x, 1.0, z);
569     PetscCallExternal(mkl_sparse_x_mv, SPARSE_OPERATION_NON_TRANSPOSE, 1.0, aijmkl->csrA, aijmkl->descr, x, 0.0, z);
632   Mat_SeqAIJMKL     *aijmkl = (Mat_SeqAIJMKL *)A->spptr;
657   if (!aijmkl->sparse_optimized || aijmkl->state != state) PetscCall(MatSeqAIJMKL_create_mkl_handle(A));
663     PetscCallExternal(mkl_sparse_x_mv, SPARSE_OPERATION_TRANSPOSE, 1.0, aijmkl->csrA, aijmkl->descr, x, 1.0, z);
667     PetscCallExternal(mkl_sparse_x_mv, SPARSE_OPERATION_TRANSPOSE, 1.0, aijmkl->csrA, aijmkl->descr, x, 0.0, z);
983   Mat_SeqAIJMKL *aijmkl;
993   PetscCall(PetscNew(&aijmkl));
994   B->spptr = (void *)aijmkl;
1002   aijmkl->sparse_optimized = PETSC_FALSE;
1003   aijmkl->no_SpMV2         = PetscDefined(HAVE_MKL_SPARSE_OPTIMIZE) ? PETSC_FALSE : PETSC_TRUE; /* Default to using the SpMV2 routines if our MKL supports them. */
1004   aijmkl->eager_inspection = PETSC_FALSE;
1008   PetscCall(PetscOptionsBool("-mat_aijmkl_no_spmv2", "Disable use of inspector-executor (SpMV 2) routines", "None", (PetscBool)aijmkl->no_SpMV2, (PetscBool *)&aijmkl->no_SpMV2, &set));
1009   PetscCall(PetscOptionsBool("-mat_aijmkl_eager_inspection", "Run inspection at matrix assembly time, instead of waiting until needed by an operation", "None", (PetscBool)aijmkl->eager_inspection, (PetscBool *)&aijmkl->eager_inspection, &set));
1012   if (!aijmkl->no_SpMV2) {
1014     aijmkl->no_SpMV2 = PETSC_TRUE;
1042   if (aijmkl->no_SpMV2) {