xref: /petsc/src/mat/impls/aij/mpi/mpicusparse/mpicusparsematimpl.h (revision d71ae5a4db6382e7f06317b8d368875286fe9008)
1519f805aSKarl Rupp #if !defined(__MPICUSPARSEMATIMPL)
2bbf3fe20SPaul Mullowney   #define __MPICUSPARSEMATIMPL
3bbf3fe20SPaul Mullowney 
4b06137fdSPaul Mullowney   #include <cusparse_v2.h>
5303a667bSSatish Balay   #include <petsc/private/cudavecimpl.h>
6b06137fdSPaul Mullowney 
7219fbbafSJunchao Zhang struct Mat_MPIAIJCUSPARSE {
8bbf3fe20SPaul Mullowney   /* The following are used by GPU capabilities to store matrix storage formats on the device */
9e057df02SPaul Mullowney   MatCUSPARSEStorageFormat   diagGPUMatFormat;
10e057df02SPaul Mullowney   MatCUSPARSEStorageFormat   offdiagGPUMatFormat;
11042217e8SBarry Smith   PetscSplitCSRDataStructure deviceMat;
12ddea5d60SJunchao Zhang   PetscInt                   coo_nd, coo_no; /* number of nonzero entries in coo for the diag/offdiag part */
13ddea5d60SJunchao Zhang   THRUSTINTARRAY            *coo_p;          /* the permutation array that partitions the coo array into diag/offdiag parts */
14ddea5d60SJunchao Zhang   THRUSTARRAY               *coo_pw;         /* the work array that stores the partitioned coo scalar values */
15219fbbafSJunchao Zhang 
16219fbbafSJunchao Zhang   /* Extended COO stuff */
17158ec288SJunchao Zhang   PetscCount  *Ajmap1_d, *Aperm1_d;            /* Local entries to diag */
18158ec288SJunchao Zhang   PetscCount  *Bjmap1_d, *Bperm1_d;            /* Local entries to offdiag */
19219fbbafSJunchao Zhang   PetscCount  *Aimap2_d, *Ajmap2_d, *Aperm2_d; /* Remote entries to diag */
20219fbbafSJunchao Zhang   PetscCount  *Bimap2_d, *Bjmap2_d, *Bperm2_d; /* Remote entries to offdiag */
21219fbbafSJunchao Zhang   PetscCount  *Cperm1_d;                       /* Permutation to fill send buffer. 'C' for communication */
22219fbbafSJunchao Zhang   PetscScalar *sendbuf_d, *recvbuf_d;          /* Buffers for remote values in MatSetValuesCOO() */
23219fbbafSJunchao Zhang   PetscBool    use_extended_coo;
24219fbbafSJunchao Zhang 
25*d71ae5a4SJacob Faibussowitsch   Mat_MPIAIJCUSPARSE()
26*d71ae5a4SJacob Faibussowitsch   {
27219fbbafSJunchao Zhang     diagGPUMatFormat    = MAT_CUSPARSE_CSR;
28219fbbafSJunchao Zhang     offdiagGPUMatFormat = MAT_CUSPARSE_CSR;
29219fbbafSJunchao Zhang     coo_p               = NULL;
30219fbbafSJunchao Zhang     coo_pw              = NULL;
31219fbbafSJunchao Zhang     deviceMat           = NULL;
32219fbbafSJunchao Zhang     use_extended_coo    = PETSC_FALSE;
33219fbbafSJunchao Zhang   }
34219fbbafSJunchao Zhang };
35bbf3fe20SPaul Mullowney #endif
36